Attached. Please let me know if you have any questions or need revisions.

A Berkeley View of Systems Challenges for AI
Ion Stoica, Dawn Song, Raluca Ada Popa, David Patterson, Michael W. Mahoney, Randy Katz,
Anthony D. Joseph, Michael Jordan, Joseph M. Hellerstein, Joseph Gonzalez, Ken Goldberg,
Ali Ghodsi, David Culler, Pieter Abbeel∗

arXiv:1712.05855v1 [cs.AI] 15 Dec 2017

ABSTRACT
With the increasing commoditization of computer vision, speech
recognition and machine translation systems and the widespread
deployment of learning-based back-end technologies such as digital advertising and intelligent infrastructures, AI (Artificial Intelligence) has moved from research labs to production. These
changes have been made possible by unprecedented levels of data
and computation, by methodological advances in machine learning,
by innovations in systems software and architectures, and by the
broad accessibility of these technologies.
The next generation of AI systems promises to accelerate these
developments and increasingly impact our lives via frequent interactions and making (often mission-critical) decisions on our behalf,
often in highly personalized contexts. Realizing this promise, however, raises daunting challenges. In particular, we need AI systems
that make timely and safe decisions in unpredictable environments,
that are robust against sophisticated adversaries, and that can process ever increasing amounts of data across organizations and individuals without compromising confidentiality. These challenges
will be exacerbated by the end of the Moore’s Law, which will constrain the amount of data these technologies can store and process.
In this paper, we propose several open research directions in systems, architectures, and security that can address these challenges
and help unlock AI’s potential to improve lives and society.

KEYWORDS
AI, Machine Learning, Systems, Security

1

foster new industries around IoT, augmented reality, biotechnology
and autonomous vehicles.
These applications will require AI systems to interact with the
real world by making automatic decisions. Examples include autonomous drones, robotic surgery, medical diagnosis and treatment,
virtual assistants, and many more. As the real world is continually changing, sometimes unexpectedly, these applications need to
support continual or life-long learning [96, 109] and never-ending
learning [76]. Life-long learning systems aim at solving multiple
tasks sequentially by efficiently transferring and utilizing knowledge from already learned tasks to new tasks while minimizing
the effect of catastrophic forgetting [71]. Never-ending learning is
concerned with mastering a set of tasks in each iteration, where
the set keeps growing and the performance on all the tasks in the
set keeps improving from iteration to iteration.
Meeting these requirements raises daunting challenges, such
as active exploration in dynamic environments, secure and robust
decision-making in the presence of adversaries or noisy and unforeseen inputs, the ability to explain decisions, and new modular
architectures that simplify building such applications. Furthermore,
as Moore’s Law is ending, one can no longer count on the rapid
increase of computation and storage to solve the problems of nextgeneration AI systems.
Solving these challenges will require synergistic innovations
in architecture, software, and algorithms. Rather than addressing
specific AI algorithms and techniques, this paper examines the
essential role that systems will play in addressing challenges in AI
and proposes several promising research directions on that frontier.

INTRODUCTION

Conceived in the early 1960’s with the vision of emulating human
intelligence, AI has evolved towards a broadly applicable engineering discipline in which algorithms and data are brought together
to solve a variety of pattern recognition, learning, and decisionmaking problems. Increasingly, AI intersects with other engineering and scientific fields and cuts across many disciplines in computing.
In particular, computer systems have already proved essential in
catalyzing recent progress in AI. Advances in parallel hardware [31,
58, 90] and scalable software systems [32, 46, 114] have sparked the
development of new machine learning frameworks [14, 31, 98] and
algorithms [18, 56, 62, 91] to allow AI to address large-scale, realworld problems. Rapidly decreasing storage costs [1, 80], crowdsourcing, mobile applications, internet of things (IoT), and the competitive advantage of data [40] have driven further investment in
data-processing systems and AI technologies [87]. The overall effect
is that AI-based solutions are beginning to approach or even surpass
human-level capabilities in a range of real-world tasks. Maturing AI
technologies are not only powering existing industries—including
web search, high-speed trading and commerce—but are helping to

2

WHAT IS BEHIND AI’S RECENT SUCCESS

The remarkable progress in AI has been made possible by a “perfect
storm” emerging over the past two decades, bringing together:
(1) massive amounts of data, (2) scalable computer and software
systems, and (3) the broad accessibility of these technologies. These
trends have allowed core AI algorithms and architectures, such as
deep learning, reinforcement learning, and Bayesian inference to
be explored in problem domains of unprecedented scale and scope.

2.1

Big data

With the widespread adoption of online global services, mobile
smartphones, and GPS by the end of 1990s, internet companies
such as Google, Amazon, Microsoft, and Yahoo! began to amass
huge amounts of data in the form of audio, video, text, and user
logs. When combined with machine learning algorithms, these
massive data sets led to qualitatively better results in a wide range of
core services, including classical problems in information retrieval,
information extraction, and advertising [49].

A Berkeley View of Systems Challenges for AI,

2.2

Big systems

Processing this deluge of data spurred rapid innovations in computer and software systems. To store massive amounts of data, internet service companies began to build massive-scale datacenters,
some of which host nearly 100, 000 servers, and provide EB [65] of
storage. To process this data, companies built new large-scale software systems able to run on clusters of cheap commodity servers.
Google developed MapReduce [32] and Google File System [43], followed shortly by the open-source counterpart, Apache Hadoop [7].
Then came a plethora of systems [46, 55, 60, 67, 114], that aimed to
improve speed, scale, and ease of use. These hardware and software
innovations led to the datacenter becoming the new computer [11].
With the growing demand for machine learning (ML), researchers and practitioners built libraries on top of these systems
to satisfy this demand [8, 52, 75].
The recent successes of deep learning (DL) have spurred a new
wave of specialized software systems have emerged to scale out
these workloads on CPU clusters and take advantage of specialized hardware, such as GPUs and TPUs. Examples include TensorFlow [2], Caffe [57], Chainer [20], PyTorch [89], and MXNet [22].

2.3

Accessibility to state-of-the-art technology

The vast majority of systems that process data and support AI workloads are built as open-source software, including Spark [114], TensorFlow [2], MXNet [22], Caffe [57], PyTorch [89], and BigDL [15].
Open source allows organizations and individuals alike to leverage
state-of-the-art software technology without incurring the prohibitive costs of development from scratch or licensing fees.
The wide availability of public cloud services (e.g., AWS, Google
Cloud, and MS Azure) allows everyone to access virtually unlimited amounts of processing and storage without needing to build
large datacenters. Now, researchers can test their algorithms at a
moment’s notice on numerous GPUs or FPGAs by spending just a
few thousands of dollars, which was unthinkable a decade ago.

3

TRENDS AND CHALLENGES

While AI has already begun to transform many application domains,
looking forward, we expect that AI will power a much wider range
of services, from health care to transportation, manufacturing to
defense, entertainment to energy, and agriculture to retail. Moreover, while large-scale systems and ML frameworks have already
played a pivotal role in the recent success of AI, looking forward,
we expect that, together with security and hardware architectures,
systems will play an even more important role in enabling the broad
adoption of AI. To realize this promise, however, we need to address
significant challenges that are driven by the following trends.

3.1

Mission-critical AI

With ongoing advances in AI in applications, from banking to
autonomous driving to robot-assisted surgery and to home automation, AI is poised to drive more and more mission-critical
applications where human well-being and lives are at stake.
As AI will increasingly be deployed in dynamic environments,
AI systems will need to continually adapt and learn new “skills”
as the environment changes. For example, a self-driving car could
quickly adapt to unexpected and dangerous road conditions (e.g.,

an accident or oil on the road), by learning in real time from other
cars that have successfully dealt with these conditions. Similarly,
an AI-powered intrusion-detection system must quickly identify
and learn new attack patterns as they happen. In addition, such
mission-critical applications must handle noisy inputs and defend
against malicious actors.
Challenges: Design AI systems that learn continually by interacting with a dynamic environment, while making decisions that are
timely, robust, and secure.

3.2

Personalized AI

From virtual assistants to self-driving cars and political campaigns,
user-specific decisions that take into account user behavior (e.g., a
virtual assistant learning a user’s accent) and preferences (e.g., a
self-driving system learning the level of “aggressiveness” a user is
comfortable with) are increasingly the focus. While such personalized systems and services provide new functionality and significant
economic benefits, they require collecting vast quantities of sensitive personal information and their misuse could affect users’
economic and psychological wellbeing.
Challenges: Design AI systems that enable personalized applications and services yet do not compromise users’ privacy and security.

3.3

AI across organizations

Companies are increasingly leveraging third-party data to augment
their AI-powered services [27]. Examples include hospitals sharing data to prevent epidemic outbreaks and financial institutions
sharing data to improve their fraud-detection capabilities. The proliferation of such applications will lead to a transition from data
silos—where one company collects data, processes it, and provides
the service—to data ecosystems, where applications learns and
make decisions using data owned by different organizations.
Challenges: Design AI systems that can train on datasets owned
by different organizations without compromising their confidentiality,
and in the process provide AI capabilities that span the boundaries of
potentially competing organization.

3.4

AI demands outpacing the Moore’s Law

The ability to process and store huge amounts of data has been one
of the key enablers of the AI’s recent successes (see Section 2.1).
However, keeping up with the data being generated will become
increasingly difficult due to the following two trends.
First, data continues to grow exponentially. A 2015 Cisco white
paper [25] claims that the amount of data generated by Internet of
Everything (IoE) devices by 2018 to be 400ZB, which is almost 50x
the estimated traffic in 2015. According to a recent study [100], by
2025, we will need a three-to-four orders of magnitude improvement in compute throughput to process the aggregate output of all
genome sequencers in the world. This would require computation
resources to at least double every year.
Second, this explosion of data is coming at a time when our
historically rapidly improving hardware technology is coming to a
grinding halt [53]. The capacity of DRAMs and disks are expected
to double just once in the next decade, and it will take two decades
before the performance of CPUs doubles. This slowdown means that
storing and processing all generated data will become impracticable.

A Berkeley View of Systems Challenges for AI
Challenges: Develop domain-specific architectures and software
systems to address the performance needs of future AI applications
in the post-Moore’s Law era, including custom chips for AI workloads, edge-cloud systems to efficiently process data at the edge, and
techniques for abstracting and sampling data.

4

RESEARCH OPPORTUNITIES

This section discusses the previous challenges from the systems
perspective. In particular, we discuss how innovations in systems,
security, and architectures can help address these challenges. We
present nine research opportunities (from R1 to R9), organized
into three topics: acting in dynamic environments, secure AI, and
AI-specific architectures. Figure 1 shows the most common relationships between trends, on one hand, and challenges and research
topics, on the other hand.
Trends
Mission-critical AI

R 1, R

2, R 3

5
,R
R4
9
,R
R7

Personalized AI

Challenges & Research

R4,

,R

R7

R5,

R6

8

AI across
organizations

AI demands outpacing
Moore’s Law

R6
R8
,R
R7, R8,

9
R9

Acting in dynamic
environments:
R1: Continual learning
R2: Robust decisions
R3: Explainable decisions
Secure AI:
R4: Secure enclaves
R5: Adversarial learning
R6: Shared learning on
confidential data
AI-specific architectures:
R7: Domain specific hardware
R8: Composable AI systems
R9: Cloud-edge systems

Figure 1: A mapping from trends to challenges and research topics.

4.1

Acting in dynamic environments

Many future AI applications will operate in dynamic environments,
i.e., environments that may change, often rapidly and unexpectedly,
and often in non-reproducible ways. For example, consider a group
of robots providing security for an office building. When one robot
breaks or a new one is added, the other robots must update their
strategies for navigation, planning, and control in a coordinated
manner. Similarly, when the environment changes, either due to the
robots’ own actions or to external conditions (e.g., an elevator going
out of service, or a malicious intruder), all robots must re-calibrate
their actions in light of the change. Handling such environments
will require AI systems that can react quickly and safely even to
scenarios that have not been encountered before.
R1: Continual learning. Most of today’s AI systems, including
movie recommendation, image recognition, and language translation, perform training offline and then make predictions online.
That is, the learning performed by the system does not happen
continually with the generation of the data, but instead it happens
sporadicallly, on very different and much slower time scales. Typically, models are updated daily, or in the best case hourly, while
predictions/decisions happen at second or sub-second granularity.
This makes them a poor fit for environments that change continually and unexpectedly, especially in mission-critical applications.

A Berkeley View of Systems Challenges for AI,
These more challenging environments require agents that continually learn and adapt to asynchronous changes.
Some aspects of learning in dynamic environments are addressed
by online learning [17], in which data arrive temporally and updates
to the model can occur as new data arrive. However, traditional
online learning does not aim to handle control problems, in which
an agent’s actions change the environment (e.g., as arise naturally
in robotics), nor does it aim to handle cases in which the outcomes
of decisions are delayed (e.g., a move in a game of chess whose
outcome is only evaluated at the end, when the game is lost or won).
These more general situations can be addressed in the framework of Reinforcement Learning (RL). The central task of RL is
to learn a function—a “policy”—that maps observations (e.g., car’s
camera inputs or user’s requested content) to actions (e.g., slowing
down the car or presenting an ad) in a sequence that maximizes
long-term reward (e.g., avoiding collisions or increasing sales). RL
algorithms update the policy by taking into account the impact of
agent’s actions on the environment, even when delayed. If environmental changes lead to reward changes, RL updates the policy
accordingly. RL has a long-standing tradition, with classical success
stories including learning to play backgammon at level of the best
human players [108], learning to walk [105], and learning basic
motor skills [86]. However, these early efforts require significant
tuning for each application. Recent efforts are combining deep
neural networks with RL (Deep RL) to develop more robust training algorithms that can work for a variety of environments (e.g.,
many Atari games [77]), or even across different application domains, as in the control of (simulated) robots [92] and the learning
of robotic manipulation skills [66]. Noteworthy recent results also
include Google’s AlphaGo beating the Go world champion [95],
and new applications in medical diagnosis [104] and resource management [33].
However, despite these successes, RL has yet to see widescale
real-world application. There are many reasons for this, one of
which is that large-scale systems have not been built with these use
cases in mind. We believe that the combination of ongoing advances
in RL algorithms, when coupled with innovations in systems design,
will catalyze rapid progress in RL and drive new RL applications.
Systems for RL. Many existing RL applications, such as gameplaying, rely heavily on simulations, often requiring millions or
even billions of simulations to explore the solution space and “solve”
complex tasks. Examples include playing different variants of a
game or experimenting with different control strategies in a robot
simulator. These simulations can take as little as a few milliseconds,
and their durations can be highly variable (e.g., it might take a
few moves to lose a game vs. hundreds of moves to win one). Finally, real-world deployments of RL systems need to process inputs
from a variety of sensors that observe the environment’s state, and
this must be accomplished under stringent time constraints. Thus,
we need systems that can handle arbitrary dynamic task graphs,
where tasks are heterogeneous in time, computation, and resource
demands. Given the short duration of the simulations, to fully utilize a large cluster, we need to execute millions of simulations per
second. None of the existing systems satisfies these requirements.
Data parallel systems [55, 79, 114] handle orders of magnitude fewer
tasks per sec, while HPC and distributed DL systems [2, 23, 82]

A Berkeley View of Systems Challenges for AI,
have limited support for heterogeneous and dynamic task graphs.
Hence, we need new systems to support effectively RL applications.
Simulated reality (SR). The ability to interact with the environment is fundamental to RL’s success. Unfortunately, in real-world
applications, direct interaction can be slow (e.g., on the order of seconds) and/or hazardous (e.g., risking irreversible physical damage),
both of which conflict with the need for having millions of interactions before a reasonable policy is learned. While algorithmic
approaches have been proposed to reduce the number of real-world
interactions needed to learn policies [99, 111, 112], more generally
there is a need for Simulated Reality (SR) architectures, in which an
agent can continually simulate and predict the outcome of the next
action before actually taking it [101].
SR enables an agent to learn not only much faster but also much
more safely. Consider a robot cleaning an environment that encounters an object it has not seen before, e.g., a new cellphone. The robot
could physically experiment with the cellphone to determine how
to grasp it, but this may require a long time and might damage the
phone. In contrast, the robot could scan the 3D shape of the phone
into a simulator, perform a few physical experiments to determine
rigidity, texture, and weight distribution, and then use SR to learn
how to successfully grasp it without damage.
Importantly, SR is quite different from virtual reality (VR);
while VR focuses on simulating a hypothetical environment (e.g.,
Minecraft), sometimes incorporating past snapshots of the real
world (e.g., Flight Simulator), SR focuses on continually simulating
the physical world with which the agent is interacting. SR is
also different from augmented reality (AR), which is primarily
concerned with overlaying virtual objects onto real world images.
Arguably the biggest systems challenges associated with SR are
to infer continually the simulator parameters in a changing realworld environment and at the same time to run many simulations
before taking a single real-time action. As the learning algorithm
interacts with the world, it gains more knowledge which can be
used to improve the simulation. Meanwhile, many potential simulations would need to be run between the agent’s actions, using
both different potential plans and making different “what-if” assumptions about the world. Thus, the simulation is required to run
much faster than real time.
Research: (1) Build systems for RL that fully exploit parallelism,
while allowing dynamic task graphs, providing millisecond-level latencies, and running on heterogeneous hardware under stringent deadlines. (2) Build systems that can faithfully simulate the real-world
environment, as the environment changes continually and unexpectedly, and run faster than real time.
R2: Robust decisions. As AI applications are increasingly
making decisions on behalf of humans, notably in mission-critical
applications, an important criterion is that they need to be robust to
uncertainty and errors in inputs and feedback. While noise-resilient
and robust learning is a core topic in statistics and machine learning,
adding system support can significantly improve classical methods.
In particular, by building systems that track data provenance, we
can diminish uncertainty regarding the mapping of data sources to
observations, as well as their impact on states and rewards. We can
also track and leverage contextual information that informs the design of source-specific noise models (e.g., occluded cameras). These
capabilities require support for provenance and noise modeling in

data storage systems. While some of these challenges apply more
generally, two notions of robustness that are particularly important
in the context of AI systems and that present particular systems
challenges are: (1) robust learning in the presence of noisy and adversarial feedback, and (2) robust decision-making in the presence
of unforeseen and adversarial inputs.
Increasingly, learning systems leverage data collected from unreliable sources, possibly with inaccurate labels, and in some cases
with deliberately inaccurate labels. For example, the Microsoft Tay
chatbot relied heavily on human interaction to develop rich natural
dialogue capabilities. However, when exposed to Twitter messages,
Tay quickly took on a dark personality [16].
In addition to dealing with noisy feedback, another research
challenge is handling inputs for which the system was never trained.
In particular, one often wishes to detect whether a query input is
drawn from a substantially different distribution than the training
data, and then take safe actions in those cases. An example of a safe
action in a self-driving car may be to slow down and stop. More
generally, if there is a human in the loop, a decision system could
relinquish control to a human operator. Explicitly training models
to decline to make predictions for which they are not confident,
or to adopt a default safe course of actions, and building systems
that chain such models together can both reduce computational
overhead and deliver more accurate and reliable predictions.
Research: (1) Build fine grained provenance support into AI systems to connect outcome changes (e.g., reward or state) to the data
sources that caused these changes, and automatically learn causal,
source-specific noise models. (2) Design API and language support for
developing systems that maintain confidence intervals for decisionmaking, and in particular can flag unforeseen inputs.
R3: Explainable decisions. In addition to making black-box
predictions and decisions, AI systems will often need to provide
explanations for their decisions that are meaningful to humans.
This is especially important for applications in which there are
substantial regulatory requirements as well as in applications such
as security and healthcare where legal issues arise [24]. Here, explainable should be distinguished from interpretable, which is often
also of interest. Typically, the latter means that the output of the
AI algorithm is understandable to a subject matter expert in terms
of concepts from the domain from which the data are drawn [69],
while the former means that one can identify the properties of the
input to the AI algorithm that are responsible for the particular
output, and can answer counterfactual or “what-if” questions. For
example, one may wish to know what features of a particular organ in an X-ray (e.g., size, color, position, form) led to a particular
diagnosis and how the diagnosis would change under minor perturbations of those features. Relatedly, one may wish to explore what
other mechanisms could have led to the same outcomes, and the
relative plausibility of those outcomes. Often this will require not
merely providing an explanation for a decision, but also considering
other data that could be brought to bear. Here we are in the domain
of causal inference, a field which will be essential in many future AI
applications, and one which has natural connections to diagnostics
and provenance ideas in databases.
Indeed, one ingredient for supporting explainable decisions is the
ability to record and faithfully replay the computations that led to a
particular decision. Such systems hold the potential to help improve

A Berkeley View of Systems Challenges for AI
decision explainability by replaying a prediction task against past
inputs—or randomly or adversarially perturbed versions of past
inputs, or more general counterfactual scenarios—to identify what
features of the input have caused a particular decision. For example,
to identify the cause of a false alarm in a video-based security
system, one might introduce perturbations in the input video that
attenuate the alarm signal (e.g., by masking regions of the image) or
search for closely related historical data (e.g., by identifying related
inputs) that led to similar decisions. Such systems could also lead
to improved statistical diagnostics and improved training/testing
for new models; e.g., by designing models that are (or are not)
amenable to explainability.
Research: Build AI systems that can support interactive diagnostic
analysis, that faithfully replay past executions, and that can help to
determine the features of the input that are responsible for a particular
decision, possibly by replaying the decision task against past perturbed
inputs. More generally, provide systems support for causal inference.

4.2

Secure AI

Security is a large topic, many aspects of which will be central to
AI applications going forward. For example, mission-critical AI
applications, personalized learning, and learning across multiple
organizations all require systems with strong security properties.
While there is a wide range of security issues, here we focus on two
broad categories. The first category is an attacker compromising
the integrity of the decision process. The attacker can do so either
by compromising and taking the control of the AI system itself, or
by altering the inputs so that the system will unknowingly render
decisions that the attacker wants. The second category is an attacker
learning the confidential data on which an AI system was trained
on, or learning the secret model. Next, we discuss three promising
research topics to defend against such attacks.
R4: Secure enclaves. The rapid rise of public cloud and the
increased complexity of the software stack considerably widen the
exposure of AI applications to attacks. Two decades ago most applications ran on top of a commercial OS, such as Windows or
SunOS, on a single server deployed behind organization’s firewalls.
Today, organizations run AI applications in the public cloud on
a distributed set of servers they do not control, possibly shared
with competitors, on a considerably more complex software stack,
where the OS itself runs on top of a hypervisor or within a container. Furthermore, the applications leverage directly or indirectly
a plethora of other systems, such as log ingestion, storage, and data
processing frameworks. If any of these software components is
compromised, the AI applications itself might be compromised.
A general approach to deal with these attacks is providing a “secure enclave” abstraction—a secure execution environment—which
protects the application running within the enclave from malicious
code running outside the enclave. One recent example is Intel’s
Software Guard Extensions (SGX) [5], which provides a hardwareenforced isolated execution environment. Code inside SGX can
compute on data, while even a compromised operating system or
hypervisor (running outside the enclave) cannot see this code or
data. SGX also provides remote attestation [6], a protocol enabling a
remote client to verify that the enclave is running the expected code.
ARM’s TrustZone is another example of a hardware enclave. At

A Berkeley View of Systems Challenges for AI,
the other end of the spectrum, cloud providers are starting to offer
special bare-bone instances that are physically protected, e.g., they
are deployed in secure “vaults” to which only authorized personnel,
authenticated via fingerprint or iris scanning, has access.
In general, with any enclave technology, the application developer must trust all the software running within the enclave. Indeed,
even in the case of hardware enclaves, if the code running inside the
enclave is compromised, it can leak decrypted data or compromise
decisions. Since a small code base is typically easier to secure, one
research challenge is to split the AI system’s code into code running
inside the enclave, hopefully as little as possible, and code running
outside of the enclave, in untrusted mode, by leveraging cryptographic techniques. Another approach to ensure that code inside
the enclave does not leak sensitive information is to develop static
and dynamic verification tools as well as sandboxing [9, 12, 93].
Note that beside minimizing the trusted computing base, there
are two additional reasons for splitting the application code: increased functionality and reduced cost. First, some of the functionality might not be available within the enclave, e.g., GPU processing
for running Deep Learning (DL) algorithms, or services and applications which are not vetted/ported yet to run within the secure
enclave. Second, the secure instances offered by a cloud provider
can be significantly more expensive than regular instances.
Research: Build AI systems that leverage secure enclaves to ensure
data confidentiality, user privacy and decision integrity, possibly by
splitting the AI system’s code between a minimal code base running
within the enclave, and code running outside the enclave. Ensure
the code inside the enclave does not leak information, or compromise
decision integrity.
R5: Adversarial learning. The adaptive nature of ML algorithms opens the learning systems to new categories of attacks that
aim to compromise the integrity of their decisions by maliciously
altering training data or decision input. There are two broad types
of attacks: evasion attacks and data poisoning attacks.
Evasion attacks happen at the inference stage, where an adversary attempts to craft data that is incorrectly classified by the
learning system [47, 103]. An example is altering the image of a
stop sign slightly such that it still appears to a human to be a stop
sign but is seen by an autonomous vehicle as a yield sign.
Data poisoning attacks happen at the training stage, where an
adversary injects poisoned data (e.g., data with wrong labels) into
the training data set that cause the learning system to learn the
wrong model, such that the adversary thereby has input data incorrectly classified by the learner [73, 74, 113]. Learning systems
that are periodically retrained to handle non-stationary input data
are particularly vulnerable to this attack, if the weakly labeled data
being used for retraining is collected from unreliable or untrustworthy sources. With new AI systems continually learning by
interacting with dynamic environments, handling data poisoning
attacks becomes increasingly important.
Today, there are no effective solutions to protect against evasion
attacks. As such, there are a number of open research challenges:
provide better understanding of why adversarial examples are often
easy to find, investigate what method or combination of different
methods may be effective at defending against adversarial examples,
and design and develop systematic methods to evaluate potential
defenses. For data poisoning attacks, open challenges include how

A Berkeley View of Systems Challenges for AI,
to detect poisoned input data and how to build learning systems
that are resilient to different types of data poisoning attacks. In
addition, as data sources are identified to be fraudulent or explicitly
retracted for regulatory reasons, we can leverage replay (see R3:
Explainable decisions) and incremental computation to efficiently
eliminate the impact of those sources on learned models. As pointed
out previously, this ability is enabled by combining modeling with
provenance and efficient computation in data storage systems.
Research: Build AI systems that are robust against adversarial
inputs both during training and prediction (e.g., decision making),
possibly by designing new machine learning models and network
architectures, leveraging provenance to track down fraudulent data
sources, and replaying to redo decisions after eliminating the fraudulent sources.
R6: Shared learning on confidential data. Today, each company typically collects data individually, analyzes it, and uses this
data to implement new features and products. However, not all
organizations possess the same wealth of data as found in the few
large AI-focused corporations, such as Google, Facebook, Microsoft,
and Amazon. Going forward, we expect more and more organizations to collect valuable data, more third-party data services to be
available, and more benefit to be gained from learning over data
from multiple organizations (see Section 3).
Indeed, from our own interaction with industry, we are learning
about an increasing number of such scenarios. A large bank provided us with a scenario in which they and other banks would like
to pool together their data and use shared learning to improve their
collective fraud detection algorithms. While these banks are natural
competitors in providing financial services, such ”cooperation” is
critical to minimize their losses due to fraudulent activities. A very
large health provider described a similar scenario in which competing hospitals would like to share data to train a shared model
predicting flu outbreaks without sharing the data for other purposes.
This would allow them to improve the response to epidemics and
contain the outbreaks, e.g., by rapidly deploying mobile vaccination
vans at critical locations. At the same time, every hospital must
protect the confidentiality of their own patients.
The key challenge of shared learning is how to learn a model
on data belonging to different (possible competitive) organizations
without leaking relevant information about this data during the
training process. One possible solution would be to pool all the
data in a hardware enclave and then learn the model. However,
this solution is not always feasible as hardware enclaves are not
yet deployed widely, and, in some cases, the data cannot be moved
due to regulatory constraints or its large volume.
Another promising approach is using secure multi-party computation (MPC) [13, 45, 70]. MPC enables n parties, each having a
private input, to compute a joint function over the input without
any party learning the inputs of the other parties. Unfortunately,
while MPC is effective for simple computations, it has a nontrivial
overhead for complex computations, such as model training. An
interesting research direction is how to partition model training
into (1) local computation and (2) computation using MPC, so that
we minimize the complexity of the MPC computation.
While training a model without compromising data confidentiality is a big step towards enabling shared learning, unfortunately,
it is not always enough. Model serving—the inferences (decisions)

rendered based on the model—can still leak information about the
data [42, 94]. One approach to address this challenge is differential
privacy [36, 37, 39], a popular technique proposed in the context of
statistical databases. Differential privacy adds noise to each query
to protect the data privacy, hence effectively trading accuracy for
privacy [35]. A central concept of differential privacy is the privacy
budget which caps the number of queries given a privacy guarantee.
There are three interesting research directions when applying
differential privacy to model serving. First, a promising approach
is to leverage differential privacy for complex models and inferences, by taking advantage of the inherent statistical nature of
the models and predictions. Second, despite the large volume of
theoretical research, there are few practical differential privacy
systems in use today. An important research direction is to build
tools and systems to make it easy to enable differential privacy for
real-world applications, including intelligently selecting which privacy mechanisms to use for a given application and automatically
converting non-differentially-private computations to differentiallyprivate computations. Finally, one particular aspect in the context
of continual learning is that data privacy can be time dependent,
that is, the privacy of fresh data is far more important than the
privacy of historical data. Examples are stock market and online
bidding, where the privacy of the fresh data is paramount, while
the historical data is sometimes publicly released. This aspect could
enable the development of new differential privacy systems with
adaptive privacy budgets that apply only to decisions on the most
recent data. Another research direction is to further develop the
notion of differential privacy under continuous observation and
data release [21, 38].
Even if we are able to protect data confidentiality during training
and decision making, this might still not be enough. Indeed, even
if confidentiality is guaranteed, an organization might refuse to
share its data for improving a model from which its competitors
might benefit. Thus, we need to go beyond guaranteeing confidentiality and provide incentives to organizations to share their
data or byproducts of their data. Specifically, we need to develop
approaches that ensure that by sharing data, an organization gets
strictly better service (i.e., better decisions) than not sharing data.
This requires ascertaining the quality of the data providing by a
given organization—a problem which can be tackled via a “leaveone-out” approach in which performance is compared both with
and without that organization’s data included in the training set.
We then provide decisions that are corrupted by noise at a level
that is inversely proportional to the quality of the data provided
by an organization. This incentivizes an organization to provide
higher-quality data. Overall, such incentives will need to be placed
within a framework of mechanism design to allow organizations to
forge their individual data-sharing strategies.
Research: Build AI systems that (1) can learn across multiple
data sources without leaking information from a data source during
training or serving, and (2) provide incentives to potentially competing
organizations to share their data or models.

4.3

AI-specific architectures

AI demands will drive innovations both in systems and hardware
architectures. These new architectures will aim not only to improve

A Berkeley View of Systems Challenges for AI
the performance, but to simplify the development of the next generation of AI applications by providing rich libraries of modules
that are easily composable.
R7: Domain specific hardware. The ability to process and
store huge amounts of data has been one of the key enablers of the
AI’s recent successes (see Section 2.1). However, continuing to keep
up with the data being generated will be increasingly challenging.
As discussed in Section 3, while data continues to grow exponentially, the corresponding performance-cost-energy improvements
that have fueled the computer industry for more than 40 years are
reaching the end-of-line:
• Transistors are not getting much smaller due to the ending
of Moore’s Law,
• Power is limiting what can be put on a chip due to the end
of Dennard scaling,
• We’ve already switched from one inefficient processor/chip
to about a dozen efficient processors per chip, but there
are limits to parallelism due to Amdahl’s Law.
The one path left to continue the improvements in performanceenergy-cost of processors is developing domain-specific processors.
These processors do only a few tasks, but they do them extremely
well. Thus, the rapid improvements in processing that we have
expected in the Moore’s law era must now come through innovations in computer architecture instead of semiconductor process
improvements. Future servers will have much more heterogeneous
processors than in the past. One trailblazing example that spotlights domain specific processors is Google’s Tensor Processing
Unit, which has been deployed in its datacenters since 2015 and
is regularly used by billions of people. It performs the inference
phase of deep neural networks 15× to 30× faster than its contemporary CPUs and GPUs and its performance per watt is 30× to
80× better. In addition, Microsoft has announced the availability
of FPGA-powered instances on its Azure cloud [88], and a host of
companies, ranging from Intel to IBM, and to startups like Cerebras and Graphcore are developing specialized hardware for AI
that promise orders of magnitude performance improvements over
today’s state-of-the-art processors [19, 48, 54, 78].
With DRAM subject to the same limitations, there are several
novel technologies being developed that hope to be its successor.
3D XPoint from Intel and Micron aims to provide 10× storage capacity with DRAM-like performance. STT MRAM aims to succeed
Flash, which may hit similar scaling limits as DRAM. Hence, the
memory and storage of the cloud will likely have more levels in the
hierarchy and contain a wider variety of technologies. Given the
increasing diversity of processors, memories, and storage devices,
mapping services to hardware resources will become an even more
challenging problem. These dramatic changes suggest building
cloud computing from a much more flexible building block than the
classic standard rack containing a top-of-rack switch and tens of
servers, each with 2 CPU chips, 1 TB of DRAM, and 4 TBs of flash.
For example, the UC Berkeley Firebox project [41] proposes a
multi-rack supercomputer that connects thousands of processor
chips with thousands of DRAM chips and nonvolatile storage chips
using fiber optics to provide low-latency, high-bandwidth, and long
physical distance. Such a hardware system would allow system
software to provision computation services with the right ratio

A Berkeley View of Systems Challenges for AI,
and type of domain-specific processors, DRAM, and NVRAM. Such
resource disaggregation at scale would significantly improve the
allocation of increasingly diverse tasks to correspondingly heterogeneous resources. It is particularly valuable for AI workloads,
which are known to gain significant performance benefits from
large memory and have diverse resource requirements that don’t
all conform to a common pattern.
Besides performance improvements, new hardware architectures
will also provide additional functionality, such as security support.
While Intel’s SGX and ARM’s TrustZone are paving the way towards hardware enclaves, much more needs to be done before they
can be fully embraced by AI applications. In particular, existing enclaves exhibit various resource limitations such as addressable memory, and they are only available for a few general purpose CPUs.
Removing these limitations, and providing a uniform hardware
enclave abstraction across a diverse set of specialized processors,
including GPUs and TPUs, are promising directions of research. In
addition, open instruction set processors, such as RISC-V represent
an exciting “playground” to develop new security features.
Research: (1) Design domain-specific hardware architectures to
improve the performance and reduce power consumption of AI applications by orders of magnitude, or enhance the security of these
applications. (2) Design AI software systems to take advantage of these
domain-specific architectures, resource disaggregation architectures,
and future non-volatile storage technologies.
R8: Composable AI systems. Modularity and composition
have played a fundamental role in the rapid progress of software
systems, as they allowed developers to rapidly build and evolve new
systems from existing components. Examples range from microkernel OSes [3, 68], LAMP stack [64], microservice architectures [85],
and the internet [26]. In contrast, today’s AI systems are monolithic
which makes them hard to develop, test, and evolve.
Similarly, modularity and composition will be key to increasing
development speed and adoption of AI, by making it easier to
integrate AI in complex systems. Next, we discuss several research
problems in the context of model and action composition.
Model composition is critical to the development of more flexible
and powerful AI systems. Composing multiple models and querying
them in different patterns enables a tradeoff between decision accuracy, latency, and throughput in a model serving system [29, 106]
In one example, we can query models serially, where each model
either renders the decision with sufficiently high accuracy or says
“I’m not sure”. In the latter case, the decision is passed to the next
model in the series. By ordering the models from the highest to the
lowest “I’m not sure” rate, and from lowest to the highest latency,
we can optimize both latency and accuracy.
To fully enable model composition, many challenges remain to
be addressed. Examples are (1) designing a declarative language
to capture the topology of these components and specifying performance targets of the applications, (2) providing accurate performance models for each component, including resource demands,
latency and throughput, and (3) scheduling and optimization algorithms to compute the execution plan across components, and
map components to the available resources to satisfy latency and
throughput requirements while minimizing costs.
Action composition consists of aggregating sequences of basic
decisions/actions into coarse-grained primitives, also called options.

A Berkeley View of Systems Challenges for AI,
In the case of a self-driving car, an example of an option is changing
the lane while driving on a highway, while the actions are speeding
up, slowing down, turning left or right, signaling the change of
direction, etc. In the case of a robot, an example of a primitive could
be grasping an object, while actions include actuating the robot’s
joints. Options have been extensively studied in the context of hierarchical learning [30, 34, 84, 97, 102, 110]. Options can dramatically
speed up learning or adaptation to a new scenario by allowing the
agent to select from a list of existing options to accomplish a given
task, rather than from a much longer list of low-level actions.
A rich library of options would enable the development of new
AI applications by simply composing the appropriate options the
same way web programmers develop applications today in just
a few lines of code by invoking powerful web APIs. In addition,
options can improve responsiveness as selecting the next action
within an option is a much simpler task than selecting an action in
the original action space.
Research: Design AI systems and APIs that allow the composition
of models and actions in a modular and flexible manner, and develop
rich libraries of models and options using these APIs to dramatically
simplify the development of AI applications.
R9: Cloud-edge systems. Today, many AI applications such
as speech recognition and language translation are deployed in the
cloud. Going forward we expect a rapid increase in AI systems that
span edge devices and the cloud. On one hand, AI systems which
are currently cloud only, such as user recommendation systems [72],
are moving some of their functionality to edge devices to improve
security, privacy, latency and safety (including the ability to cope
with being disconnected from the internet). On the other hand,
AI systems currently deployed at the edge, such as self-driving
cars, drones, and home robots, are increasingly sharing data and
leveraging the computational resources available in the cloud to
update models and policies [61].
However, developing cloud and the cloud-edge systems is challenging for several reasons. First, there is a large discrepancy between the capabilities of edge devices and datacenter servers. We
expect this discrepancy to increase in the future, as edge devices,
such as cellphones and tablets, have much more stringent power and
size constraints than servers in datacenters. Second, edge devices
are extremely heterogeneous both in terms of resource capabilities,
ranging from very low power ARM or RISC-V CPUs that power IoT
devices to powerful GPUs in self-driving cars, and software platforms. This heterogeneity makes application development much
harder. Third, the hardware and software update cycles of edge
devices are significantly slower than in a datacenter. Fourth, as the
increase in the storage capacity slows down, while the growth in
the data being generated continues unabated, it may no longer be
feasible or cost effective to store this deluge of data.
There are two general approaches to addressing the mix of cloud
and edge devices. The first is to repurpose code to multiple heterogeneous platforms via retargetable software design and compiler
technology. To address the wide heterogeneity of edge devices
and the relative difficulty of upgrading the applications running on
these devices, we need new software stacks that abstract away the
heterogeneity of devices by exposing the hardware capabilities to
the application through common APIs. Another promising direction is developing compilers and just-in-time (JIT) technologies to

efficiently compile on-the-fly complex algorithms and run them on
edge devices. This approach can leverage recent code generation
tools, such as TensorFlow’s XLA [107], Halide [50], and Weld [83].
The second general approach is to design AI systems that are
well suited to partitioned execution across the cloud and the edge.
As one example, model composition (see Section 4.3) could allow
one to run the lighter but less accurate models at the edge, and the
computation-intensive but higher-accuracy models in the cloud.
This architecture would improve decision latency, without compromising accuracy, and it has been already employed in recent video
recognition systems [59, 115]. In another example, action composition would allow building systems where learning of hierarchical
options [63] takes place on powerful clusters in the cloud, and then
execution of these options happens at the edge.
Robotics is one application domain that can take advantage of a
modular cloud-edge architecture. Today, there is a scarcity of open
source platforms to develop robotic applications. ROS, arguably
the most popular such platform in use today, is confined to running locally and lacks many performance optimizations required
by real-time applications. To take advantage of the new developments in AI research such as shared and continual learning, we
need systems that can span both edge devices (e.g., robots) and
the cloud. Such systems would allow developers to seamlessly migrate the functionality between a robot and the cloud to optimize
decision latency and learning convergence. While the cloud can
run sophisticated algorithms to continually update the models by
leveraging the information gathered by distributed robots in real
time, the robots can continue to execute the actions locally based
on previously downloaded policies.
To address the challenge of the data deluge collected by the edge
devices, learning-friendly compression methods can be used to
reduce processing overhead. Examples of such methods include
sampling and sketching, which have already been successfully employed for analytics workloads [4, 10, 28, 51, 81]. One research
direction is to aggressively leverage sampling and sketching in a
systematic way to support a variety of learning algorithms and prediction scenarios. An arguably more difficult challenge is to reduce
the storage overhead, which might require to delete data. The key
challenge here is that we do not always know how the data will be
used in the future. This is essentially a compression problem, but
compression for the purposes of ML algorithms. Again, distributed
approaches based in materialized samples and sketches can help
provide solutions to this problem, as can ML-based approaches in
the form of feature selection or model selection protocols.
Research: Design cloud-edge AI systems that (1) leverage the edge
to reduce latency, improve safety and security, and implement intelligent data retention techniques, and (2) leverage the cloud to share
data and models across edge devices, train sophisticated computationintensive models, and take high quality decisions.

5

CONCLUSION

The striking progress of AI during just the last decade is leading
to the successful transition from the research lab into commercial
services that have previously required human input and oversight.
Rather than replacing human workers, AI systems and robots have

A Berkeley View of Systems Challenges for AI
potential to enhance human performance and facilitate new forms
of collaboration [44].
To realize the full promise of AI as a positive force in our lives,
there are daunting challenges to overcome, and many of these challenges are related to systems and infrastructure. These challenges
are driven by the realization that AI systems will need to make
decisions that are faster, safer, and more explainable, securing these
decisions as well as the learning processes against ever more sophisticated types of attacks, continuously increasing the computation
capabilities in the face of the end of Moore’s Law, and building composable systems that are easy to integrate in existing applications
and can span the cloud and the edge.
This paper proposes several open research directions in systems,
architectures, and security that have potential to address these
challenges. We hope these questions will inspire new research that
can advance AI and make it more capable, understandable, secure
and reliable.

REFERENCES
[1] A History of Storage Cost. 2017.
http://www.mkomo.com/
cost-per-gigabyte-update. (2017).
[2] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen,
Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, and Matthieu Devin.
2015. TensorFlow: Large-scale machine learning on heterogeneous systems.
(2015).
[3] Mike Accetta, Robert Baron, William Bolosky, David Golub, Richard Rashid,
Avadis Tevanian, and Michael Young. 1986. Mach: A New Kernel Foundation for
UNIX Development. 93–112.
[4] Sameer Agarwal et al. 2013. BlinkDB: queries with bounded errors and bounded
response times on very large data. In EuroSys.
[5] Ittai Anati, Shay Gueron, Simon Johnson, and Vincent Scarlata. 2013. Innovative
technology for CPU based attestation and sealing. In Proceedings of the 2nd
international workshop on hardware and architectural support for security and
privacy, Vol. 13.
[6] Ittai Anati, Shay Gueron, Simon P Johnson, and Vincent R Scarlata. 2013. Innovative Technology for CPU Based Attestation and Sealing. (2013).
[7] Apache Hadoop. 2017. http://hadoop.apache.org/. (2017).
[8] Apache Mahout. 2017. http://mahout.apache.org/. (2017).
[9] Sergei Arnautov, Bohdan Trach, Franz Gregor, Thomas Knauth, Andre Martin,
Christian Priebe, Joshua Lind, Divya Muthukumaran, Daniel OfiKeeffe, Mark L
Stillwell, et al. 2016. SCONE: Secure linux containers with Intel SGX. In 12th
USENIX Symp. Operating Systems Design and Implementation.
[10] Peter Bailis, Edward Gan, Samuel Madden, Deepak Narayanan, Kexin Rong, and
Sahaana Suri. 2017. MacroBase: Prioritizing Attention in Fast Data. In Proceedings
of the 2017 ACM International Conference on Management of Data (SIGMOD ’17).
ACM, New York, NY, USA, 541–556.
[11] Luiz Andre Barroso and Urs Hoelzle. 2009. The Datacenter As a Computer: An
Introduction to the Design of Warehouse-Scale Machines. Morgan and Claypool.
[12] Andrew Baumann, Marcus Peinado, and Galen Hunt. 2015. Shielding applications
from an untrusted cloud with haven. ACM Transactions on Computer Systems
(TOCS) 33, 3 (2015), 8.
[13] Michael Ben-Or, Shafi Goldwasser, and Avi Wigderson. 1988. Completeness
theorems for non-cryptographic fault-tolerant distributed computation. In Proceedings of the 20th ACM symposium on Theory of Computing.
[14] James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan
Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua
Bengio. 2010. Theano: a CPU and GPU math expression compiler. In Proceedings
of the Python for scientific computing conference (SciPy), Vol. 4. Austin, TX, 3.
[15] bigdl. BigDL: Distributed Deep Learning on Apache Spark. https://software.intel.
com/en-us/articles/bigdl-distributed-deep-learning-on-apache-spark. (????).
[16] Tay (bot). 2017. https://en.wikipedia.org/wiki/Tay (bot). (2017).
[17] Léon Bottou. 1998. On-line Learning in Neural Networks. (1998), 9–42.
[18] Léon Bottou. 2010. Large-scale machine learning with stochastic gradient descent.
In Proceedings of COMPSTAT’2010. Springer, 177–186.
[19] Cerebras. 2017. https://www.cerebras.net/. (2017).
[20] Chainer. 2017. https://chainer.org/. (2017).
[21] T.-H. Hubert Chan, Elaine Shi, and Dawn Song. 2010. Private and Continual
Release of Statistics. In ICALP (2), Vol. 6199. Springer.
[22] Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun
Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. 2015. MXNet: A Flexible
and Efficient Machine Learning Library for Heterogeneous Distributed Systems.

A Berkeley View of Systems Challenges for AI,
arXiv preprint arXiv:1512.01274 (2015).
[23] Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun
Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. 2015. MXNet: A Flexible
and Efficient Machine Learning Library for Heterogeneous Distributed Systems.
CoRR abs/1512.01274 (2015).
[24] Travers Ching, Daniel S Himmelstein, Brett K Beaulieu-Jones, Alexandr A Kalinin,
Brian T Do, Gregory P Way, Enrico Ferrero, Paul-Michael Agapow, Wei Xie,
Gail L Rosen, et al. 2017. Opportunities And Obstacles For Deep Learning In
Biology And Medicine. bioRxiv (2017), 142760.
[25] cisco. 2015. Cisco Global Cloud Index: Forecast and Methodology, 20152020. http://www.cisco.com/c/dam/en/us/solutions/collateral/service-provider/
global-cloud-index-gci/white-paper-c11-738085.pdf. (2015).
[26] D. Clark. 1988. The Design Philosophy of the DARPA Internet Protocols. SIGCOMM Comput. Commun. Rev. 18, 4 (Aug. 1988), 106–114.
[27] CMS updates rule allowing claims data to be sold. 2016. http://www.
modernhealthcare.com/article/20160701/NEWS/160709998. (2016).
[28] Graham Cormode, Minos Garofalakis, Peter J Haas, and Chris Jermaine. 2012.
Synopses for massive data: Samples, histograms, wavelets, sketches. Foundations
and Trends in Databases 4, 1–3 (2012), 1–294.
[29] Daniel Crankshaw, Xin Wang, Giulio Zhou, Michael J. Franklin, Joseph E. Gonzalez, and Ion Stoica. 2017. Clipper: A Low-Latency Online Prediction Serving
System. NSDI ’17 (2017).
[30] Peter Dayan and Geoffrey E. Hinton. 1992. Feudal Reinforcement Learning. In
Advances in Neural Information Processing Systems 5, [NIPS Conference, Denver,
Colorado, USA, November 30 - December 3, 1992]. 271–278. http://papers.nips.cc/
paper/714-feudal-reinforcement-learning
[31] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark
Mao, Marcaurelio Ranzato, Andrew Senior, Paul Tucker, Ke Yang, Quoc Le,
and Andrew Y. Ng. 2012. Large Scale Distributed Deep Networks. In NIPS ’12.
http://papers.nips.cc/paper/4687-large-scale-distributed-deep-networks.pdf
[32] Jeffrey Dean and Sanjay Ghemawat. 2004. MapReduce: Simplified Data Processing on Large Clusters. In Proceedings of the 6th Conference on Symposium on
Opearting Systems Design & Implementation - Volume 6 (OSDI’04).
[33] DeepMind
AI
Reduces
Google
Data
Centre
Cooling
Bill
by
40%.
2017.
https://deepmind.com/blog/
deepmind-ai-reduces-google-data-centre-cooling-bill-40/. (2017).
[34] Thomas G. Dietterich. 1998. The MAXQ Method for Hierarchical Reinforcement
Learning. In Proceedings of the Fifteenth International Conference on Machine
Learning (ICML 1998), Madison, Wisconsin, USA, July 24-27, 1998. 118–126.
[35] John Duchi, Michael Jordan, and Martin Wainwright. to appear. Minimax optimal
procedures for locally private estimation. J. Amer. Statist. Assoc. (to appear).
[36] Cynthia Dwork. 2006. Differential Privacy. In ICALP (2), Vol. 4052. Springer.
[37] Cynthia Dwork. 2008. Differential privacy: A survey of results. In International
Conference on Theory and Applications of Models of Computation.
[38] Cynthia Dwork, Moni Naor, Toniann Pitassi, and Guy N Rothblum. 2010. Differential privacy under continual observation. In Proceedings of the 42nd ACM
symposium on Theory of computing.
[39] Cynthia Dwork and Aaron Roth. 2014. The algorithmic foundations of differential
privacy. Foundations and Trends in Theoretical Computer Science 9 (2014).
[40] The Economist. 2017. The world’s most valuable resource is no longer oil, but
data. (May 2017).
[41] FireBox. 2017. https://bar.eecs.berkeley.edu/projects/2015-firebox.html. (2017).
[42] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. 2015. Model inversion
attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications
Security. ACM, 1322–1333.
[43] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung. 2003. The Google File
System. In Proceedings of the Nineteenth ACM Symposium on Operating Systems
Principles (SOSP ’03). 29–43.
[44] Ken Goldberg. 2017. Op-Ed: Call it Multiplicity: Diverse Groups of People and
Machines Working Together. Wall Street Journal (2017).
[45] Oded Goldreich, Silvio Micali, and Avi Wigderson. 1987. How to play any mental
game. In Proceedings of the 19th ACM symposium on Theory of computing.
[46] Joseph E. Gonzalez, Yucheng Low, Haijie Gu, Danny Bickson, and Carlos Guestrin.
2012. PowerGraph: Distributed Graph-parallel Computation on Natural Graphs
(OSDI’12). 17–30.
[47] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and
harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014).
[48] Graohcore. 2017. https://www.graphcore.ai/. (2017).
[49] Alon Halevy, Peter Norvig, , and Fernando Pereira. 2009. The Unreasonable
Effectiveness of Data. IEEE Intelligent Systems 24, 2 (2009), 8–12.
[50] Halide: A Language for Image Processing and Computational Photography. 2017.
http://halide-lang.org/. (2017).
[51] Joseph M. Hellerstein, Peter J. Haas, and Helen J. Wang. 1997. Online Aggregation.
In Proceedings of the 1997 ACM SIGMOD International Conference on Management
of Data (SIGMOD ’97). ACM, New York, NY, USA, 171–182. https://doi.org/10.
1145/253260.253291

A Berkeley View of Systems Challenges for AI,
[52] Joseph M. Hellerstein, Christoper Ré, Florian Schoppmann, Daisy Zhe Wang,
Eugene Fratkin, Aleksander Gorajek, Kee Siong Ng, Caleb Welton, Xixuan Feng,
Kun Li, and Arun Kumar. 2012. The MADlib Analytics Library: Or MAD Skills,
the SQL. Proc. VLDB Endow. 5, 12 (Aug. 2012), 1700–1711.
[53] John L. Hennessy and David A. Patterson. to appear. Computer Architecture,
Sixth Edition: A Quantitative Approach. (to appear).
[54] Intel Nervana. 2017. https://www.intelnervana.com/intel-nervana-hardware/.
(2017).
[55] Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, and Dennis Fetterly. 2007.
Dryad: Distributed Data-parallel Programs from Sequential Building Blocks. In
Proceedings of the 2nd ACM SIGOPS/EuroSys European Conference on Computer
Systems 2007 (EuroSys ’07). 59–72.
[56] Martin Jaggi, Virginia Smith, Martin Takac, Jonathan Terhorst, Sanjay Krishnan, Thomas Hoffmann, and Michael I. Jordan. 2015. Communication-Efficient
Distributed Dual Coordinate Ascent. In NIPS, 27.
[57] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long,
Ross Girshick, Sergio Guadarrama, and Trevor Darrell. 2014. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the ACM
International Conference on Multimedia. ACM, 675–678.
[58] Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal,
Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle,
Pierre-luc Cantin, Clifford Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt
Dau, Jeffrey Dean, Ben Gelb, Tara Vazir Ghaemmaghami, Rajendra Gottipati,
William Gulland, Robert Hagmann, C. Richard Ho, Doug Hogberg, John Hu,
Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander
Kaplan, Harshit Khaitan, Daniel Killebrew, Andy Koch, Naveen Kumar, Steve
Lacy, James Laudon, James Law, Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle
Lucke, Alan Lundin, Gordon MacKean, Adriana Maggiore, Maire Mahony, Kieran
Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy Nix, Thomas Norrie,
Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Matt Ross,
Amir Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham,
Jed Souter, Dan Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo
Tian, Horia Toma, Erick Tuttle, Vijay Vasudevan, Richard Walter, Walter Wang,
Eric Wilcox, and Doe Hyun Yoon. 2017. In-Datacenter Performance Analysis
of a Tensor Processing Unit. In Proceedings of the 44th Annual International
Symposium on Computer Architecture (ISCA ’17). ACM, New York, NY, USA, 1–12.
https://doi.org/10.1145/3079856.3080246
[59] Daniel Kang, John Emmons, Firas Abuzaid, Peter Bailis, and Matei Zaharia.
2017. Optimizing Deep CNN-Based Queries over Video Streams at Scale. CoRR
abs/1703.02529 (2017).
[60] Asterios Katsifodimos and Sebastian Schelter. 2016. Apache Flink: Stream Analytics at Scale.
[61] Ben Kehoe, Sachin Patil, Pieter Abbeel, and Ken Goldberg. 2015. A Survey of
Research on Cloud Robotics and Automation. IEEE Trans. Automation Science
and Eng. 12, 2 (2015).
[62] Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic
Optimization. CoRR abs/1412.6980 (2014). http://arxiv.org/abs/1412.6980
[63] Sanjay Krishnan, Roy Fox, Ion Stoica, and Ken Goldberg. 2017. DDCO: Discovery
of Deep Continuous Options for Robot Learning from Demonstrations. In 1st
Conference on Robot Learning (CoRL).
[64] LAMP (software bundle). 2017. https://en.wikipedia.org/wiki/LAMP (software
bundle). (2017).
[65] Leo Leung. 2015. How much data does x store? (March 2015). https:
//techexpectations.org/tag/how-much-data-does-youtube-store/
[66] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. 2016. End-to-end
Training of Deep Visuomotor Policies. J. Mach. Learn. Res. 17, 1 (Jan. 2016),
1334–1373. http://dl.acm.org/citation.cfm?id=2946645.2946684
[67] Mu Li, David G. Andersen, Jun Woo Park, Alexander J. Smola, Amr Ahmed,
Vanja Josifovski, James Long, Eugene J. Shekita, and Bor-Yiing Su. 2014. Scaling
Distributed Machine Learning with the Parameter Server. In OSDI ’14. 583–598.
[68] J. Liedtke. 1995. On Micro-kernel Construction. In Proceedings of the Fifteenth
ACM Symposium on Operating Systems Principles (SOSP ’95). ACM, New York,
NY, USA, 237–250. https://doi.org/10.1145/224056.224075
[69] M. W. Mahoney and P. Drineas. 2009. CUR Matrix Decompositions for Improved
Data Analysis. Proc. Natl. Acad. Sci. USA 106 (2009), 697–702.
[70] Dahlia Malkhi, Noam Nisan, Benny Pinkas, Yaron Sella, et al. 2004. FairplaySecure Two-Party Computation System.. In USENIX Security Symposium, Vol. 4.
San Diego, CA, USA.
[71] Michael Mccloskey and Neil J. Cohen. 1989. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. The Psychology of Learning
and Motivation 24 (1989), 104–169.
[72] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and
Blaise Aguera y Arcas. 2016. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference
on Artificial Intelligence and Statistics (AISTATS). http://arxiv.org/abs/1602.05629
[73] Shike Mei and Xiaojin Zhu. 2015. The Security of Latent Dirichlet Allocation.. In
AISTATS.

[74] Shike Mei and Xiaojin Zhu. 2015. Using Machine Teaching to Identify Optimal
Training-Set Attacks on Machine Learners.. In AAAI. 2871–2877.
[75] Xiangrui Meng, Joseph Bradley, Burak Yavuz, Evan Sparks, Shivaram Venkataraman, Davies Liu, Jeremy Freeman, DB Tsai, Manish Amde, Sean Owen, Doris
Xin, Reynold Xin, Michael J. Franklin, Reza Zadeh, Matei Zaharia, and Ameet
Talwalkar. 2016. MLlib: Machine Learning in Apache Spark. Journal of Machine
Learning Research 17, 34 (2016), 1–7. http://jmlr.org/papers/v17/15-237.html
[76] Tom M Mitchell, William W Cohen, Estevam R Hruschka Jr, Partha Pratim
Talukdar, Justin Betteridge, Andrew Carlson, Bhavana Dalvi Mishra, Matthew
Gardner, Bryan Kisiel, Jayant Krishnamurthy, et al. 2015. Never Ending Learning..
In AAAI. 2302–2310.
[77] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness,
Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg
Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen
King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. 2015.
Human-level control through deep reinforcement learning. Nature 518, 7540 (26
02 2015), 529–533. http://dx.doi.org/10.1038/nature14236
[78] Dharmendra Modha. 2016.
The brainfis architecture, efficiencyfi on
a chip.
(Dec. 2016).
https://www.ibm.com/blogs/research/2016/12/
the-brains-architecture-efficiency-on-a-chip/
[79] Derek G. Murray, Malte Schwarzkopf, Christopher Smowton, Steven Smith, Anil
Madhavapeddy, and Steven Hand. 2011. CIEL: A Universal Execution Engine for
Distributed Data-flow Computing. In Proceedings of the 8th USENIX Conference
on Networked Systems Design and Implementation (NSDI’11). USENIX Association,
Berkeley, CA, USA, 113–126. http://dl.acm.org/citation.cfm?id=1972457.1972470
[80] Average Historic Price of RAM. 2017. http://www.statisticbrain.com/averagehistoric-price-of-ram/. (2017).
[81] Frank Olken and Doron Rotem. 1990. Random sampling from database files: A
survey. Statistical and Scientific Database Management (1990), 92–111.
[82] Open MPI: Open Source High Performance Computing. 2017. https://www.
open-mpi.org/. (2017).
[83] Shoumik Palkar, James J. Thomas, Anil Shanbhag, Deepak Narayanan, Holger
Pirk, Malte Schwarzkopf, Saman Amarasinghe, and Matei Zaharia. 2017. Weld:
A Common Runtime for High Performance Data Analytics. In CIDR.
[84] Ronald Parr and Stuart J. Russell. 1997. Reinforcement Learning with Hierarchies
of Machines. In Advances in Neural Information Processing Systems 10, [NIPS
Conference, Denver, Colorado, USA, 1997]. 1043–1049. http://papers.nips.cc/paper/
1384-reinforcement-learning-with-hierarchies-of-machines
[85] Pattern: Microservice Architecture. 2017. http://microservices.io/patterns/
microservices.html. (2017).
[86] Jan Peters and Stefan Schaal. 2008. Reinforcement learning of motor skills with
policy gradients. Neural networks 21, 4 (2008), 682–697.
[87] Gil Press. 2016. Forrester Predicts Investment In Artificial Intelligence Will Grow
300% in 2017. Forbes (November 2016).
[88] Project Catapult. 2017. https://www.microsoft.com/en-us/research/project/
project-catapult/. (2017).
[89] PyTorch. 2017. http://pytorch.org/. (2017).
[90] Rajat Raina, Anand Madhavan, and Andrew Y. Ng. 2009. Large-scale Deep
Unsupervised Learning Using Graphics Processors. In Proceedings of the 26th
Annual International Conference on Machine Learning (ICML ’09). ACM, New
York, NY, USA, 873–880. https://doi.org/10.1145/1553374.1553486
[91] Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. 2011. Hogwild:
A Lock-Free Approach to Parallelizing Stochastic Gradient Descent. In NIPS 24.
[92] John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter
Abbeel. 2015. Trust Region Policy Optimization. In Proceedings of the 32nd
International Conference on Machine Learning (ICML).
[93] Felix Schuster, Manuel Costa, Cédric Fournet, Christos Gkantsidis, Marcus
Peinado, Gloria Mainar-Ruiz, and Mark Russinovich. 2015. VC3: Trustworthy data analytics in the cloud using SGX. In Security and Privacy (SP), 2015 IEEE
Symposium on. IEEE, 38–54.
[94] Reza Shokri, Marco Stronati, and Vitaly Shmatikov. 2016. Membership inference
attacks against machine learning models. arXiv preprint arXiv:1610.05820 (2016).
[95] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George
Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, and others. 2016. Mastering the game of Go with deep neural
networks and tree search. Nature 529, 7587 (2016), 484–489.
[96] Daniel L Silver, Qiang Yang, and Lianghao Li. 2013. Lifelong Machine Learning
Systems: Beyond Learning Algorithms.. In AAAI Spring Symposium: Lifelong
Machine Learning, Vol. 13. 05.
[97] Satinder P. Singh. 1992. Reinforcement Learning with a Hierarchy of Abstract
Models. In Proceedings of the 10th National Conference on Artificial Intelligence.
San Jose, CA, July 12-16, 1992. 202–207. http://www.aaai.org/Library/AAAI/1992/
aaai92-032.php
[98] Evan R Sparks et al. 2013. MLI: An API for distributed machine learning. In
ICDM.
[99] Stephane Ross. 2013. Interactive Learning for Sequential Decisions and Predictions. https://en.wikipedia.org/wiki/LAMP (software bundle). (2013).

A Berkeley View of Systems Challenges for AI
[100] Zachary D Stephens, Skylar Y Lee, Faraz Faghri, Roy H Campbell, Chengxiang
Zhai, Miles J Efron, Ravishankar Iyer, Michael C Schatz, Saurabh Sinha, and
Gene E Robinson. 2015. Big data: Astronomical or genomical? PLoS Biology 13,
7 (2015), e1002195.
[101] Richard S. Sutton. Integrated architectures for learning, planning, and reacting
based on approximating dynamic programming. In Proceedings of the Seventh
International Conference on Machine Learning. Morgan Kaufmann.
[102] Richard S. Sutton, Doina Precup, and Satinder P. Singh. 1999. Between MDPs and
Semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning.
Artif. Intell. 112, 1-2 (1999), 181–211. https://doi.org/10.1016/S0004-3702(99)
00052-1
[103] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru
Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural
networks. arXiv preprint arXiv:1312.6199 (2013).
[104] Kai-Fu Tang, Hao-Cheng Kao, Chun-Nan Chou, and Edward Y. Chang. 2016.
Inquire and Diagnose: Neural Symptom Checking Ensemble using Deep Reinforcement Learning. http://infolab.stanford.edu/∼echang/NIPS DeepRL 2016
Symptom Checker.pdf. (2016).
[105] Russ Tedrake, Teresa Weirui Zhang, and H Sebastian Seung. 2005. Learning to
walk in 20 minutes. In Proceedings of the Fourteenth Yale Workshop on Adaptive
and Learning Systems, Vol. 95585. Yale University New Haven (CT), 1939–1412.
[106] TensorFlow Serving. 2017. https://tensorflow.github.io/serving/. (2017).
[107] TensorFlow XLA. 2017. https://www.tensorflow.org/performance/xla/. (2017).
[108] Gerald Tesauro. 1995. Temporal difference learning and TD-Gammon. Commun.
ACM 38, 3 (1995), 58–68.
[109] Sebastian Thrun. 1998. Lifelong learning algorithms. Learning to learn 8 (1998),
181–209.
[110] Sebastian Thrun and Anton Schwartz. 1994. Finding Structure in Reinforcement Learning. In Advances in Neural Information Processing Systems 7, [NIPS
Conference, Denver, Colorado, USA, 1994]. 385–392. http://papers.nips.cc/paper/
887-finding-structure-in-reinforcement-learning
[111] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba,
and Pieter Abbeel. 2017. Domain Randomization for Transferring Deep Neural
Networks from Simulation to the Real World. CoRR abs/1703.06907 (2017).
http://arxiv.org/abs/1703.06907
[112] Eric Tzeng, Judy Hoffman, Ning Zhang, Kate Saenko, and Trevor Darrell.
2014. Deep Domain Confusion: Maximizing for Domain Invariance. CoRR
abs/1412.3474 (2014). http://arxiv.org/abs/1412.3474
[113] Huang Xiao, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert,
and Fabio Roli. 2015. Is feature selection secure against training data poisoning?.
In ICML. 1689–1698.
[114] Matei Zaharia et al. 2012. Resilient distributed datasets: A fault-tolerant abstraction for in-memory cluster computing. In NSDI ’12.
[115] Haoyu Zhang, Ganesh Ananthanarayanan, Peter Bodı́k, Matthai Philipose,
Paramvir Bahl, and Michael J. Freedman. 2017. Live Video Analytics at Scale
with Approximation and Delay-Tolerance. In NSDI.

A Berkeley View of Systems Challenges for AI,


Theoretical and Applied Economics
FFet al
Volume XXI (2014), No. 11(600), pp. 43-56

The challenges of accounting profession
as generated by controlling
Costantin ROMAN
Bucharest University of Economic Studies, Romania
roman_costantin@yahoo.com
Aureliana-Geta ROMAN
Bucharest University of Economic Studies, Romania
romanaureliana@yahoo.de
Elena MEIER
Contrast Management Consulting & Training, Bucharest, Romania
elena.meier@contrast-consulting.com
Abstract. Internationalization and globalization marked in Europe in the early
’70s the development of controlling, in its role to complete the management and
have influenced it to a great extent until present. In practice, controllers’
responsibilities range from the extreme of performing only primary accounting
tasks and up to the extreme of undertakings related to executive management.
Controlling has undergone decades of evolution, and herewith the concept got
standardized in theory and practice; however it cannot be claimed that a unitary
definition for controlling has already been achieved. Same as a theatre spotlight
shines on the stage from different angles, controlling should be viewed in the same
way, as: mind-set, substance, work approach, process, tool and role. Only when all
the lights are on and the object is viewed from all sides, it can be defined in its
completeness. As a management function, controlling must be practiced in a
professional manner, as the sustainable development is ensured only for those
entities that act professionally. Without solid controlling know-how, managers are
sentenced to being led rather than leading. If the manager does not have solid
controlling know-how, he must rely on a team of specialists who have extensive
knowledge of controlling.
Keywords: modern management, management functions, professional controlling,
normative management, strategic management, project controlling.
JEL Classification: M21.
REL Classification: 14B.

44

Costantin Roman, Aureliana-Geta Roman, Elena Meier

Despite the revolution of information technology, mankind is still far from being
able to effortlessly satisfy its needs.
Information as a product of the human mind seems to defy economic theory
models, as it is “mobile” and can be transmitted without any losses.
As far as information is concerned, the consumption process is in fact adding
value by making use of its synergetic features. Information is valid for a long time
span because it seized a piece of reality. Even though information does not
disappear, from an economic perspective, its usage can be blocked. By blocking
the circuit of the potential contained by the information, the consumption cancels
its conversion into value, hence making it obsolete.
A well-founded and consumed piece of information can substitute substance,
energy and particularly, another piece of information.
The error of mistaking value for cost, “contributed” to the discriminatory
treatment of environment and of social services, as value creators (education,
healthcare, art, etc.). Unfortunately, even today, the world’s „decisions factors”
look at the expenses generated by schools, hospitals, research centers only from a
value perspective and not also from the perspective of the social potential that
these institutions generate for economy.
Business management cannot be achieved based on intuition and experience
alone. Assimilation and learning from past experiences is essential but not
sufficient; modern management is inconceivable without creativity. Creativity is a
synergistic blend between intuition and, above all, hard work; without intuition,
knowledge is lacking the means and thinking is lacking the content.
Wisdom gathered from all places and times is a unique product of civilization
centered on asserting the value of each person. Science is sentenced to develop
models with the aim of solving complex problems.
The complex phenomenon of macro-competence can be viewed as a strategic
structure of the entity upon which systemic offers generating outstanding and hard
to imitate value, can be build. Competence can be viewed as performing better
than others, and especially as having the capacity to provide explanations. If
analysed alone, competence has at any rate the following features: knowledge,
skill and existence.
In practice, the engineer's work is focused on studying and implementing projects
while the economist’s work focuses on managing scarce resources for the
production and distribution of goods and services. As resources are scarce,
engineers and economists meet in the “cost arena”. Both engineers and

The challenges of accounting profession as generated by controlling

45

economists resort to models that imply concept, empirical data and reality
validation.
Today, economists’ models resemble those of engineers, because each has to
consider systems’ steering or control, to strive to obtain cheap solutions and to
take actions which must have a strategic perspective.
The control of the resources implies ever more the control of the networks and
herewith of their main nodes, including their human dimensions. Globalization
and multiple ways of governance are attained through an overlap of powers and a
gradual erasing of the borders as radical separation lines between states.
The acceleration of technical progress and the massive increase of
interdependencies are factors which positively impact living conditions for most
people today and tomorrow, but they create countless problems which must be
clearly identified.
Interactions among national economies are a development factor for economies, at
both country and global level. Although in recent decades interdependencies have
increased, this did not diminish the role of national economies nor did it amputate
the national sovereignty, as practice has proved that each solid national economy
participates more effectively to the global exchange.
Each country, rich or poor, has something to offer and something to receive from
the global circuit of values. Any modern nation, while preserving its specific
character, its cultural originality, shows at the same time willingness to assimilate
from other nations’ civilization and cultural conquests, as these are universal and
have the ability to overcome local or national boundaries.
Knowledge is essential for any economic culture, as it enables mankind to grasp
as much as possible and to open new paths towards its hopes and desires.
Science and culture store millennial creations of human thinking and action, and
help mankind turn simultaneously into social creation and causal factor in the
evolutionary process of society.
The Institute for Business Administration of the Vienna University of Economics
and Business and Controller Institut Austria, have been doing researches and have
been providing courses in the field of controlling for more than 35 years and for
the past 20 years the Universities of Applied Sciences have joined as well,
emphasizing the practical approach and especially the advanced knowledge in
controlling.
The launch of Professional controlling. Concepts and Instruments, second edition
revised, published in 2011 by Schäffer-Poeschel Stuttgart Publishing, written by
Professor Rolf Eschenbach, PhD - Vienna University of Economics and Business,

45

46

Costantin Roman, Aureliana-Geta Roman, Elena Meier

founder of Controller Institut Austria, partner at Contrast Management Consulting
GmbH and by mag. doctor Helmut Siller, MSc - University of Applied Sciences
Vienna, was an important moment to acknowledge how indispensable and
essential this work is for the students and, in particular, for the specialists in
Romania.
To be mentioned that from the accepted version of the Romanian translation and
until the actual printing of the book, many specialists have contributed.
The translation from German and the selection of the most suitable terms in
Romanian language was not at all an easy task.
Let's see why this work it is so valuable and, especially, what is controlling!
The professional controlling is no longer the classic, traditional controlling of the
years 1960-1970. As a management function, controlling must be practiced in a
professional manner, as the sustainable development is ensured only for the
entities that act professionally. Professionalization means performing an activity,
a profession with the mind-set of meeting goals at the highest level.
Among many other prerequisites for this profession, we stress vocational training
which requires time, practical experience, scientifically grounded know-how,
mastering of specialized terminology, high social and technical skills for solving
challenges.
The authors tackle entities (companies, non-profit and public institutions) as being
systems, from the perspective of the management-oriented entity.
Using a systemic approach has the following advantages in the case of
controlling:
 the use of a language of systems with a high degree of generalization simplifies
analysis, absorbing knowledge, interdisciplinary transfer of knowledge;
 its transparent language allows the systematization of knowledge, the
separation of what is essential from nonessential, enables new solutions for
problem-solving;
 holistic thinking, specific to the systemic approach, is multifactorial and multicausal and takes into consideration retroactive, subsequent and delayed effects.
The book presents clearly how sustainable controlling can be successfully
performed, thanks to its professionalization.
Without solid controlling know-how, managers are sentenced to being led rather
than leading.

The challenges of accounting profession as generated by controlling

47

Controlling is a functional management concept with the role of coordinating the
planning, the control and the information towards the desired end-result. The
controller can be regarded as the “economic conscience” of the entity.
We must distinguish between controlling as a function in the organization and the
controller as a person holding this function. In fact, controlling, in terms of
steering the organization, is one of the central management activities. Each
manager fulfils controlling functions within its duties. Controlling, as both process
and mind-set, is thus generated by the manager and controller as a team and
becomes a type of “interface”. The connection between the manager’s
responsibilities, controlling and the controller’s responsibilities is illustrated in the
figure below.
Figure1. Controlling at the intersection between the manager and the controller
CONTROLLER

Responsible for results as a:
- project manager;
- product manager;
- line manager.
and for successful strategic positions.

Responsible for transparency as a
pilot on the road to profit throughout
the process of:
‐ providing information;
‐ decisions making;
‐ ensuring coordination as well as a
‐ moderator of the planning process.

Controlling

MANAGER

Controlling does not target a position or a person, but rather a field of activities
carried out by various employees or even managers who do not necessarily hold
the “controller” position. In small and medium-sized entities, controlling function
is taken over by the management of the entity or by the head of the financialaccounting department. Entities with over two hundred employees appoint more
and more often a controller to take over controlling tasks.
The coordination responsibilities of the controller translate into making sure that
the planning and control activities performed by management are goal-oriented
and that all necessary information is available at any time. Controller’s role
regarding planning is to coordinate the partial plans and to organize the entire
planning process. Therefore, it is not the controller who normally plans and
coordinates, but the manager. To define the limit, it has to be mentioned that in
small and medium-sized entities, it might occur that controller’s role surpasses
coordination responsibilities. Thus, in current practice, the controller often takes
the responsibility of planning, which actually should be performed by specialized
departments. In the recent years, it can be noticed that, at global level, the
controller’s role has extended from a simple service provider to a consultant for
the management.

47

48

Costantin Roman, Aureliana-Geta Roman, Elena Meier

The controller’s tasks, identity and responsibility have been established in the
controller model developed by the IGC (International Group of Controlling, former
Society of interests in the field of controlling). In the version of September 2002 of
this model, the responsibility of the controller for meeting the targets was
acknowledged for the first time. On one hand, controller’s liability results from
his/her responsibility for the accuracy of the information collected and provided; on
the other hand, it results from his/her responsibility for organizing and monitoring
the process, which enables the management to take target-oriented decisions.
Controller’s mission
Controllers design and accompany the management process of goal-finding,
planning and controlling and thus are co-responsible for reaching the objectives.
This means:
 Controllers ensure the transparency of business results, finance, processes and
strategy and thus contribute to higher economic effectiveness.
 Controllers coordinate sub-targets and sub-plans in a holistic way and
organise a reporting-system that is oriented towards the future and covers the
enterprise as a whole.
 Controllers moderate and design the process of goal-finding, planning and
management control so that every decision-maker can act in accordance with
agreed objectives.
 Controllers provide all relevant controlling-information to managers.
 Controllers develop and maintain the controlling systems.
The coordination activity of controlling consists in solving problems that have a
strong effect over the entity as a result of environmental influences:
 the dynamics is increasing;
 the markets are stagnating;
 new technologies are issued very fast;
 the product life cycles are becoming shorter and shorter.
Controlling helps the management of the entity to withstand these problems by
making use of innovative solutions rather than relying on old, obsolete methods.
Controlling is currently not achieved only through the controller, but often on the
spot, through directly involved employees. Controlling becomes more and more
an integrated controlling; the controlling as an institution and the controller act
more and more often as a moderator for promoting the idea of controlling.
Controlling – main elements
A careful analysis of the literature allows us to emphasize six concepts outlined
until today:

The challenges of accounting profession as generated by controlling








49

controlling as an administrative record tracking (the ’80s)
controlling as an administrative information system (end of the ’80s);
controlling as planning and control (beginning of the ’90s);
controlling as coordination activity (the ’90s);
controlling as business administration (end of the ’90s);
controlling as a system for coordinating decision-making process (the 2000s).

Within each concept, controlling is regarded differently; later concepts develop
previous ones, focusing on different aspects.
Tasks of the controlling system
According to the definition of controlling, the controller has responsibilities with
regard to planning, control and information.
In order to explain the basic idea and the need for controlling, some classic
questions might be useful, such as:
 Do you know precisely which products generate profit and from where losses
are coming?
 Do know how various measures impact the outcome?
 Do you know how your result looks like without fiscal or balance sheet
fragmentations?
 Are objective-oriented targets included in your planning and are resources
allocated properly?
 How fast can you find out if you are performing according to your plan or you
have lost control?
 Are decisions taken in due time and are proper measures put in place?
 Can you translate your company’s strategy into concrete target-oriented plans?
 Do you know the factors that generate higher indirect costs?
Each entity follows a certain strategy whose achievement is guaranteed through
proper structuring of the operational processes and through the configuration of an
appropriate organizational structure.
The most important information source within the information system is
accounting. Electronic data processing system has become an indispensable
element. Budgeting is an important component of the planning and control
processes. However, controller’s work does not cover only short-term operational
planning; taking into consideration also strategic aspects is increasingly defining
the activity of the controller.
Basically, the controller has two different coordination tasks, with regard to both
the planning system and the information system. On the one hand the controller
deals with configuration and development and on the other hand with the
permanent coordination of activities.

49

50

Costantin Roman, Aureliana-Geta Roman, Elena Meier

The purpose of controlling within the entity is to sort individual components that
generally occur, check their utility, complete and organize them within a system.
The most important components of the management system to which the
controller directs his/her activity are:
 planning and control system;
 information system.
The objective of the controlling system of the entity is to increase transparency as
a prerequisite for optimal management of the business. Controlling follows three
main principles:
 plan-oriented actions;
 decentralized, individual responsibility;
 measurable KPIs.
It is enough to remember that controller’s responsibilities and position are
determined significantly by the size of the entity. Thus, for medium-sized entities,
it is typical that the controller is seen as a “handyman” and is responsible for more
than just controlling activities, while large companies split the activities within the
financial department between the controller and the treasurer, with delimitations
that can be more or less rigid.
The job description is the most suitable instrument to formalize the framework of
the controller’s activity. Below an example of job description for a controller:
Job description
1 Job title
2 Level
Controlling coordinator
Field coordinator
3 Objectives
- development and implementation of procedures which lead to required profit for the entity
- supporting the company’s management towards efficient control from the analyst and consultant position
4 Direct superior
The CEO
4.1 Job holder receives expert guidance from
5. Job holder provides expert guidance to
- the coordinator of the cost calculation department
- the coordinator of the planning and reporting department
- the employees in the controlling department
6. Job holder is be deputised by
- the director in charge with economic, financial and administrative matters
- the coordinator of the administrative department
7. Job holder deputises
The coordinator of the administrative department
8. Special empowerment
(here must be mentioned the special powers and rights that are not specific as per the hierarchical level and that exceed
the general regulation)
- general power of attorney
- power of attorney for bank

The challenges of accounting profession as generated by controlling

51

9.Description of the specialized activities that must be carried out especially by the person holding the position
(independently)
- consultancy for the company’s CEO
- responsibility for reporting and information management systems
- development of budgets and monthly results
- deviations analysis and benchmarking
- calculation, interpretation and comments of the comparison between forecasted and actual result
- forecasting
- development of product cost calculations and target pricing
- consultancy for internal and external reporting
- calculations of profitability and investment
- financial planning
- support / perform strategic planning
- analysis of the processes and development of mitigations measure
- standardisation and development of controlling tools
- project management
Note to job holder:
The job description establishes your responsibilities and competences in a committed manner. You must act and decide
accordingly. You must inform immediately your direct superior in case of any irregularities.
Date:
Date:
Date:
I received the job description
Date:
Direct Superior
Department
Human Resources
Job holder’s signature:
coordinator
Department

Controlling in commercial entities
In the meantime, efficient controlling systems have emerged within the
commercial field as well. Among the particularities of trade is the increasingly
important role given to:
 the location or locations – for subsidiary-based systems;
 life cycles of various types of operations/commercial transactions concepts
(discounters, supermarkets, consumer markets, specialized markets, etc.);
 the range of products in its wideness and depth, against the importance of
individual item;
 the high informational and coordination needs related to developing the
product range, on behalf of the procurement and distribution department
(buying and selling);
 the ordering of the physical flow of goods;
 the organization and usage of the display areas for goods presentation;
 the coordination of the involved factors: personnel and sales spaces.
Mohlenbruch and Meier (1998) show how a unitary consistent system of
controlling for the commercial field should be in order to meet these specific
elements. At the heart of the integrated controlling system for retail entities lays a
central database that must support and improve the information transfer, as well as
the management of the interfaces.

51

52

Costantin Roman, Aureliana-Geta Roman, Elena Meier

The significant elements of the system are the following:
 overall controlling of the entity: financial controlling, strategy and organization planning, investment controlling, subsidiaries coordination, development of the types of entities, overall clients management;
 human resources controlling: flexible planning of human resources allocation
and productivity measurement based on the data from the “personnel
qualification controlling”;
 product range controlling: product range planning (inclusion or exclusion of
articles, as well as overall management of merchandise groups), product range
coordination, product range control;
 spaces/areas controlling: the management of the commercial space using the
analysis of sales data recorded by scanners, the sales integrated management of
commercial space, the management of commercial space by means of
management systems;
 goods controlling: ordering system, external logistics system, internal logistics
system, goods reception and warehouse coordination, goods delivery.
Controlling in public administration
The introduction of controlling in the public administration was triggered at the
beginning, on one hand by the high and increasing indebtedness, on the other hand by
the faulty quality of services in public administration, strongly criticized by citizens.
Compared to private companies, public administration entities do not have
services programs and the “products” are not clearly defined. The reporting, a
significant communication tool within private companies was lacking almost
entirely. Because of insufficient cost calculation, a thorough control of services is
impossible or insufficient for most positions in the public administration. One of
the reasons is that certain concepts related to information and communication
technology are, at best, in an early stage of development. Gaps are identified
mainly in the way the public administration presents itself towards its target
audience, the citizens as well as in the personnel related area.
In this context, in order to simplify administration, the so-called New Public
Management (NPM) was developed by communities in order to reform these entities.
With the help of the New Public Management, the introduction of controlling
within public administration was initiated.
The elements of the New Public Management:
 clear delimitation of responsibilities between public administration and politics;
 reduction of the centralized control with better support for decentralized areas;
 contract management;
 decentralized responsibility for resources;

The challenges of accounting profession as generated by controlling






53

development of the organization;
actions oriented towards output (costs), respectively the product;
staff training;
introduction of economy- oriented management tools (cost calculation and
profitability calculation, indicators, reporting).

The fundamentals of controlling can be very well transferred to the public
administration; however certain peculiarities need to be considered. For controlling
within the public sector, respectively for planning, management and controlling for
public administration organizations, the “3E” model (effectiveness, efficiency,
economy) according to Becker and Weise proved to be a suitable reference.
The focal point of controlling activities in terms of effective management consists
in the strategic planning and the planning of the objectives, as well as the control
of the goals achievement. Successful management is possible only when the
objectives of the civil servants are set between political and public administration
institutions involved in the structuring process and when there is a clearly
articulated outlook in terms of objectives. To meet these objectives, a series of
values under the format of indices and indicators, which should be subsequently
anchored in the public administration entities’ internal and external reporting
system, must be processed. To support the strategic planning process for
objectives and resources, as well as for the monitoring of the achievement of the
goals within the multidimensional target system of the civil servants, the Balanced
Scorecard concept has proved to be the appropriate solution, being successfully
used for many years in these entities
Figure 2. The “3E” model of public administration controlling
Goals
Resource
allocation/in
put

Planning
level
Effectiveness

Efficiency

Quality

Production
process
Cost
efficiency

Achievement
and control

Finsihed
products/output
Degree
achievement

of

Legitimacy

53

54

Costantin Roman, Aureliana-Geta Roman, Elena Meier

In terms of efficiency control, the main focus is on the planning and the control of
the productivity and profitability of the services, also taking into account the
correctitude and quality requirements of the service.
Within this operational area of controlling, we can find the annual planning and
the budgeting, the cost management and the quality management. Within the New
Public Management, a significant role in budgeting is represented by the
consistent mix between the financial responsibility and the technical (expert)
responsibility. The starting point for budgeting is represented by those expert
services which can independently decide whether to provide the services while
using more personnel or more capital. Taking a decision requires a clear concept
of what: products, product range can be made available and their subsequent cost,
as well as a proper costs and services calculation. On this basis, comparisons
between planned and actual values can be made and deviations can be analysed.
In order to achieve continuous efficiency and effectiveness, the management of
the public administration is trained to use an up-to-date, stepwise, clear and
action-oriented reporting system.
Controlling in public administration is still at the beginning of a complex
development process. In the past years, internal professionalization within public
administration followed its course.
Controlling in healthcare institutions
Healthcare is another area where controlling penetrated.
Economic planning and services control within hospitals require a service and
cost accounting which meets the following accounting objectives:
 transparency over individual activities/processes with the purpose of recording
the services delivered to each patient from admission to release;
 subsequent calculation of total costs of the services performed with the purpose
of control over the profitability against the revenues generated by the
corresponding DRGs (Diagnosis Related Groups);
 making available information in order to ensure a minimum quality for each
treatment with as little as possible consumption of resources: herein lies the
influence of medical and administrative behaviour; at the same time, the
comparison with other hospitals, respectively benchmarking should be
possible.
As long as the specificities of the hospital are considered, the process cost
calculation seems appropriate and represents a necessary prerequisite for
economic planning, for the management and control of the hospital, as well as for
negotiations with health insurance funds and authorities.

The challenges of accounting profession as generated by controlling

55

Other controlling tasks with regard to hospitals are related to conducting of
market analysis for areas of interest, analysis of doctors who issue medical
recommendations for further investigations, analysis of diagnoses and their
frequency, as well as staff analysis in order to derive constructive proposals for
future services and staff. The controller is also the moderator of the discussions
between head physicians and management with regard to hospital budget.
Recently, the term medical controlling was chosen in order to designate topics
which require special medical know-how regarding hospital management, for
example checking of the codes assigned to diagnoses and services, processing of
the treatment procedures and deviations analysis. Medical controlling implies a
close cooperation between economic and medical staff and demonstrates that in
healthcare units, controlling cannot be delegated to a single position, but it must
be fulfilled jointly by several persons.

References
Adam, D. (2000). Investitionscontrolling, 3, Auflage, München, Wien
Aichele, C. (2006). Intelligentes Projektmanagement, Stuttgart
Bähr, U. (2002). Controlling in der öffentlichen Verwaltung. Sternenfels
Baumgartner, B. (1980). Die Controller-Konzeption. Theoretische Darstellung und praktische
Anwendung, Bern, Stuttgart
Busch, V. (2004). Wettbewerbsbezogene Controllinginstrumente in Ramehn des new Public
Management, München
Deyhle, A. (2003). Controller Praxis, Band I: Unternehmensplannung, Rechnungswesen und
Controller-Funktion, 15. Auflage, Offenburg und Wörtersee
Eschenbach, R., Kreuzer, C., Kriegler-Lenz, A., Haschka, A., Kast, J. (eds.) (2007). Controlling
macht Schule. Eisenstadt
Eschenbach, R., Eschenbach, S., Kunesch, H. (2008). Strategische Konzepte- Ideen und
Instrumente von Igor Ansoff bis Hans Ulrich, 5. Auflage, Stuttgart
Hoffmann, W., Niedermayr, R., Risak, J. (1996). Führungsergänzung durch Controlling. In:
Eschenbach, R. (ed.): Controlling, 2. Auflage, Stuttgart, S. 3-48
Kotler, P., Keller, K.L., Bliemel, F. (2007). Marketing – Management, 12. Auflage, München
Küpper, H.-U. (2005). Controlling: Konzeption, Augaben, Instrumente, 4. Auflage, Stuttgart
Meier, W. (1987). Durchsetzen von Strategien – Verhaltensorientierte Führungskonzept zum
Aufbau strategischer Erfolgspositionen, 2. Auflage, Zürich
Preißler, P. R. (2000). Controlling, 12. Auflage, München, Wien
Reichmann, T. (2001). Controlling, 3. Auflage, München
Roman, C., Moşteanu, T. (2011). Finanţele instituţiilor publice, Editura Economică, București

55

56

Costantin Roman, Aureliana-Geta Roman, Elena Meier

Roman, C., Tabără, V., Roman, A.-G. (2007). Control financiar şi audit public, Editura
Economică, București
Siller, H. (1985). Grundsätze des ordnungsmäßigen strategischen Controlling. Schriften des
Österreichischen Controller-Instituts, Band I, Wien
Töpfer, A. (1976). Plannungs- und Kontrollsysteme industrieller Unternehmungen, Berlin
Weber, J. (2008). Rollen der Controller – Theoretische Herleitung und empirische Erkenntnisse.
In: Weber, J., Vater, H., Schmidt, W., Reinhard, H., Ernst, E. (2008). Die neue Rolle des
Controllers, Stuttgart, S. 3-14


Open Journal of Business and Management, 2018, 6, 850-856
http://www.scirp.org/journal/ojbm
ISSN Online: 2329-3292
ISSN Print: 2329-3284

Analysis of the Impact of Artificial Intelligence
Application on the Development of Accounting
Industry
Jiaxin Luo, Qingjun Meng, Yan Cai
School of Business, University of Hohai, Nanjing, China

How to cite this paper: Luo, J.X., Meng,
Q.J. and Cai, Y. (2018) Analysis of the
Impact of Artificial Intelligence Application
on the Development of Accounting Industry. Open Journal of Business and Management, 6, 850-856.
https://doi.org/10.4236/ojbm.2018.64063
Received: July 30, 2018
Accepted: August 28, 2018
Published: August 31, 2018
Copyright © 2018 by authors and
Scientific Research Publishing Inc.
This work is licensed under the Creative
Commons Attribution International
License (CC BY 4.0).
http://creativecommons.org/licenses/by/4.0/
Open Access

Abstract
With the rapid development of information technology and the needs of
economic society, artificial intelligence has ushered in the golden age. The
application of artificial intelligence technology in the accounting field is an
inevitable trend, which will bring tremendous changes and development to
the accounting industry. This paper takes the application of artificial intelligence in the accounting industry as the research object, analyzes the impact
of artificial intelligence on the development of accounting industry, and puts
forward relevant suggestions for its existing problems.

Keywords
Artificial Intelligence, Accounting, Transformation

1. Introduction
In 1956, at the University of Dartmouth seminar in the United States, John
McCarthy and other computer experts first proposed the concept of “artificial
intelligence” [1], marking the birth of artificial intelligence. Nowadays, with the
wave of artificial intelligence sweeping across the globe, the International Joint
Conference on Artificial Intelligence continues to research related technologies,
and the world’s major developed countries regard the development of artificial
intelligence as a major strategy to enhance national competitiveness. In 2017, artificial intelligence was first written into the Chinese government work report,
and 15 departments including the Ministry of Finance worked together to build
the world’s major artificial intelligence innovation center [2]. Nowadays, artificial intelligence technology has been widely used in agriculture, commerce,
education, and service industries. The golden age of artificial intelligence has ar-

DOI: 10.4236/ojbm.2018.64063 Aug. 31, 2018

850

Open Journal of Business and Management

J. X. Luo et al.

rived.
With the rapid development of artificial intelligence technology and its wide
application in various fields, the trend of human work being replaced by robots
is intensifying. The involvement of artificial intelligence in the accounting industry will inevitably affect and subvert the traditional development mode and
bring innovation to the accounting industry. The 2017 Chinese Government
Work Report promoted the development of artificial intelligence to the national
strategic level, clearly pointing out the need to accelerate the development and
transformation of artificial intelligence technology; the State Council issued the
“New Generation Artificial Intelligence Development Plan” to promote the integration of artificial intelligence and various industries, promote large-scale application of artificial intelligence, and comprehensively enhance the level of intelligent development of industry [3]. In February 2018, the “Standards for the
Training and Evaluation of Accounting Practice Information Application Capabilities” was approved in Beijing. Its normative content covers two sections: artificial intelligence application and management accounting information application. The support of relevant policies and the issue of regulations mean that
the combination of artificial intelligence and accounting industry is an inevitable
trend in the future.
In 2016, Deloitte Touche Tohmatsu, one of the world’s four largest accounting
firms, announced the introduction of artificial intelligence into accounting, taxation, and auditing [4]. In 2017, Nigel Duffy, a world-renowned machine learning
and artificial intelligence expert, joined Ernst & Young as the person in charge of
its global innovative artificial intelligence team. In recent years, many countries
have been competing to carry out research and application of artificial intelligence, and the call for the use of artificial intelligence is louder and louder in
academia. Professor Sun Zheng, former president of Shanghai University of
Finance and Economics, put forward the view that enterprises should use the
Internet and big data for transformation and innovation in the speech of “Financial Transformation and Evolution in the Period of VUCA”; in the keynote
speech of “New Era, New Challenges, New Changes”, Professor Qin Rongsheng,
secretary of the Party Committee of Beijing National Accounting Institute, declared that financial management should be intelligent in the future financial
management transformation of enterprises; Li Wei, an expert in China’s management accounting informationization, pointed out that the deep application of
artificial intelligence in the financial field is the general trend of financial intelligence in the future. Facing the transformation and upgrading of the accounting
industry, the traditional accounting work has the characteristics of repetition
and cumbersome. The application of artificial intelligence can solve the pain
points of inefficiency and low added value in the accounting field, making the
accountants turn to more creative work and bring greater value to the company.
In summary, the application of artificial intelligence to the accounting industry
will promote the development and innovation of the industry and enhance the
competitiveness of enterprises, which is of great significance.
DOI: 10.4236/ojbm.2018.64063

851

Open Journal of Business and Management

J. X. Luo et al.

2. Problems of the Application of Artificial Intelligence in
the Accounting Field
2.1. Lack of Experience in the Initial Stage
At present, the application of artificial intelligence in the field of accounting in
China is still in its infancy. Although more and more units have introduced accounting robots, these robots are just a kind of operation of process automation
technology with clear algorithm, which is suitable for a large number of repeated
operation scenarios without the ability of deep learning, which is essentially
“weak AI”. Currently, the application of automation technology is more extensive, and the application scope is mainly at the level of financial reporting. It has
not yet entered the core areas of accounting such as financial analysis, and has
not yet exerted an influence that can drive the change of accounting standards
[5]. When artificial intelligence is put into accounting work, it must replace
every work step of traditional accounting, including the input of original documents, the formation of accounting information, the generation of financial reports, and appropriate decision-making suggestions, so as to truly improve the
overall financial work [6]. Therefore, whether it is the depth or breadth of artificial intelligence, the application of artificial intelligence in the accounting industry is still in its infancy, and the complexity of artificial intelligence technology
and the lack of application experience have brought great difficulties to its development. Therefore, there is still a long way to go for the development of AI in
the accounting field.

2.2. High Investment with Slow Return
For enterprises, in order to introduce artificial intelligence into the accounting
field, it is necessary to design a unique artificial intelligence system that conforms to the characteristics of the enterprise according to the actual situation of
the enterprise. First of all, capital investment is the most important guarantee;
secondly, after the introduction of technology, it is necessary to adjust the management of human resources and the daily operation mode of the enterprise. Finally, when the intelligent transformation of the accounting information system
is completed, a series of training should be carried out, including the training for
the use of new system features and the training of information security [7]. Due
to the personalized features of intelligent systems, enterprises need a large
amount of resources in the initial application and later operation, which poses
great challenges to the cost control of enterprises. Given the high investment and
slow return, many enterprises may focus on short-term profits instead of making
strategic adjustments and thus stop at the introduction of artificial intelligence
technology.

2.3. The Quality of Professional Talents Needs to Be Improved
The application of artificial intelligence technology in the accounting field requires the corresponding professional talents to manage, and the current seDOI: 10.4236/ojbm.2018.64063

852

Open Journal of Business and Management

J. X. Luo et al.

nior accounting talents in China are scarce. It is estimated that as of 2015,
China’s accountants reached 20.5 million, accounting graduates with a bachelor degree or above accounted for 23.11%, and there are 100,000 certified public accountants [8]. At this stage, the structural contradiction of Chinese accounting talents is outstanding: the general basic accounting personnel are
surplus while the top accounting talents are in short supply. The combination
of artificial intelligence and accounting work puts higher demand on accountants. Accounting personnel not only need professional knowledge in accounting field, but also need to master information technology, acquire the
skillful use of accounting software and data management, so as to adapt the
changes of new work situation.

2.4. Accounting Personnel Training Program in Colleges Needs
Adjustment
Nowadays, most universities have set relevant courses in accounting computerization. However, influenced by many internal and external factors, there are
some problems such as the unitary contents of courses, the disconnection between theoretical knowledge and practical application, and the difficulty of
forming scientific computerization system, which cannot meet the requirements
of the development of the times [9]. College graduates are the main force of the
accounting field in the future. However, at present, the talent training programs
in Chinese universities fail to make corresponding adjustments in the context of
accounting reform, lack information technology courses related to artificial intelligence, and pay less attention to the innovation of accounting concepts. This
will lead to the lack of market competitiveness of college graduates and the inability to meet the market demand of the accounting industry in the future.

3. Suggestions on Improving the Effectiveness of Artificial
Intelligence Application in Accounting Field
3.1. The Government Vigorously Supports the Application of
Artificial Intelligence in the Accounting Field
With the general trend of economic globalization, the links and cooperation between countries are getting closer and closer, meanwhile, the competition is becoming more intense. To be competitive in the international stage, China’s accounting industry must constantly improve the level of the accounting industry.
At present, China has raised the development of artificial intelligence to the national strategic level. The government should also implement relevant plans and
measures in various industries and actively create favorable environmental conditions for the development of artificial intelligence in the accounting field. Relevant policies and regulations are needed to encourage and guide the application of artificial intelligence in the accounting industry. For example, enterprises
that actively apply artificial intelligence technology should be given appropriate
subsidies or tax reduction.
DOI: 10.4236/ojbm.2018.64063

853

Open Journal of Business and Management

J. X. Luo et al.

3.2. Enterprise Management Attaches Great Importance to the
Application of Artificial Intelligence Technology
At present, technology giants such as Apple, Google, and Microsoft are investing
more and more resources to seize the artificial intelligence market. The domestic
Internet leader “BAT” is also actively deploying artificial intelligence and applying this technology to more life scenarios. For example, Alibaba’s face recognition technology in payment and financial services, and intelligent robot customer service on the shopping platform, these specific applications of artificial intelligence technology have driven the transformation and upgrading of business
and services. Artificial intelligence has become a new focus for enterprises to
improve their core competitiveness. Therefore, enterprises should look upon artificial intelligence from the perspective of the overall situation, and cooperate
with cloud computing and big data analysis technology to make good use of resources. Applying artificial intelligence at the accounting level will increase the
cost of the enterprise in the short term, but looking forward to the future, it will
enable the company to have more sustainable development capabilities and will
occupy a place in the future of artificial intelligence.

3.3. Improve the Quality of Accounting Education in Colleges
Colleges and universities should arrange major courses reasonably and scientifically according to their teaching objectives, and design systematic professional
training programs. On the one hand, colleges need to attach great importance to
the combination of students’ theoretical knowledge and accounting practice,
strengthen cooperation between universities and enterprises, and strive to create
practical opportunities to improve students’ practical ability and cultivate applied accountants. On the other hand, colleges should constantly improve the
quality of teachers. Teachers should always keep abreast of the changes and
trends in the development of international financial standards and accounting in
order to make accounting classroom teaching and practical teaching keep pace
with the development of the times [10]. Only by striving to improve the teaching
level can we cultivate excellent accounting talents that meet market demands.

3.4. Accounting Talents Establish the Idea of Life-Long Learning
Accountants should pay attention to the improvement of personal professional
skills and the cultivation of professional ethics, and establish the idea of life-long
learning to adapt to the rapid changes and development requirements of the accounting industry in the future. The “13th Five-Year Plan for Accounting Reform
and Development” proposes to promote the widespread application of management accounting and implement the accounting talent strategy [11], which has
pointed out the direction for the development of accounting talents in China. At
present, China’s accounting industry is in the transitional stage from basic financial accounting to management accounting. Accounting personnel should
change their concepts, actively study and creatively apply artificial intelligence
DOI: 10.4236/ojbm.2018.64063

854

Open Journal of Business and Management

J. X. Luo et al.

and other related knowledge, and strive to become a high-level comprehensive
accounting talent, in order to avoid being eliminated by the market.

4. Conclusion
As one of the important representatives of the new round of scientific and technological revolution, artificial intelligence is moving from technology research
and development to industrial application, and has become a new driving force
for global economic development. At present, China’s economic development
has entered a new normal and the Chinese government has laid out the artificial
intelligence industry in many aspects. The accounting industry should also
strengthen the main position of artificial intelligence application in the process
of reform and innovation. For enterprises, making good use of the new information technology will be the key to capturing opportunities and upgrading in the
new era. Undoubtedly, intelligent finance and accounting is the future development trend. In the process of promoting the application of artificial intelligence
in the accounting field, it is necessary for the country, enterprises, universities,
individuals and other parties to work together, and how to effectively solve the
problems arising in the process of application will be the key.

Fund Project
2018 “Project of the National College Student Innovation and Entrepreneurship
Training Program” of Hohai University: “Analysis of the Complex Relationship
between R&D Investment and Innovation Performance of Artificial Intelligence
Companies” (No. 69).

Conflicts of Interest
The authors declare no conflicts of interest regarding the publication of this paper.

References

DOI: 10.4236/ojbm.2018.64063

[1]

Wang, Y.J. (2017) Discussion on Artificial Intelligence and Future Business Model.
China Journal of Commerce, 17 , 132-133. (In Chinese)

[2]

Xu, Z.J. (2017) 15 Ministries and Commissions Work Together to Build a New
Generation of Innovative Open Platform for Artificial Intelligence in China. Science
& Technology Industry of China, 12, 80-81. (In Chinese)

[3]

Zhang, M. (2016) Will Accountants and Tax Agents Be Replaced by Robots After
Deloitte Introduces Artificial Intelligence? Wallstreetcn.
https://wallstreetcn.com/articles/231439

[4]

State Council (2017) Notice of the State Council on Printing a New Generation of
Artificial Intelligence Development Plan. Central People’s Government of the
People’s Republic of China.
http://www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm

[5]

Chen, X.Z. (2017) “Artificial Intelligence” Storm in the Financial Field. Corporate
Finance, 10, 88-89. (In Chinese)

855

Open Journal of Business and Management

J. X. Luo et al.
[6]

Zhu, Y.Y. and Zhang, J.S. (2018) Application and Development of AI in Accounting
Industry. China Township Enterprises Accounting, 6, 264-265. (In Chinese)

[7]

Chen, K. (2017) Research on the Application of Intelligent Accounting Information
System in G Tobacco Business Enterprises. Xi’an University of Post and Telecommunications. (In Chinese)

[8]

Ministry of Finance of the People’s Republic of China (2016) Research Report on
the Supply and Demand of Accounting Talents in China. Central People’s Government of the People’s Republic of China.
http://kuaiji.firstacc.cn/a/20161022/5196.html

[9]

Hu, L.R. (2016) Analysis of Computerized Accounting Teaching Mode in Colleges
and Universities in the Internet+ Times. Modern Economic Information, 17, 271.
(In Chinese)

[10] Lian, X.L. (2006) Discussion on the Current Structural Contradiction of Accounting
Talents in China and Its Countermeasures. Contemporary Manager, 21, 942-943.
(In Chinese)
[11] Ministry of Finance of the People’s Republic of China (2016) Notice on Printing
and Distributing the Outline of the 13th Five-Year Plan for Accounting Reform and
Development. Central People’s Government of the People’s Republic of China.
http://kjs.mof.gov.cn/zhengwuxinxi/zhengcefabu/201610/t20161018_2437976.html

DOI: 10.4236/ojbm.2018.64063

856

Open Journal of Business and Management


JOURNAL OF EMERGING TECHNOLOGIES IN ACCOUNTING
Vol. 14, No. 1
Spring 2017
pp. 115–122

American Accounting Association
DOI: 10.2308/jeta-51730

The Emergence of Artificial Intelligence: How Automation is
Changing Auditing
Julia Kokina
Thomas H. Davenport
Babson College
ABSTRACT: This paper provides an overview of the emergence of artificial intelligence in accounting and auditing.
We discuss the current capabilities of cognitive technologies and the implications these technologies will have on
human auditors and the audit process itself. We also provide industry examples of artificial intelligence
implementation by Big 4 accounting firms. Finally, we address some potential biases associated with the creation
and use of artificial intelligence and discuss implications for future research.
Keywords: artificial intelligence; automation; auditing.

INTRODUCTION
Impact of Artificial Intelligence on Accounting and Auditing

T

he field of accounting in general and auditing in particular is undergoing a fundamental change due to advances in data
analytics and artificial intelligence (AI) (Agnew 2016). This paper is motivated by the need to more deeply explore the
use of artificial intelligence in accounting (Sutton, Holt, and Arnold 2016). The 2015 Deep Shift: Technology Tipping
Points and Societal Impact survey of 816 executives from the information technology and communications sector reported that
75 percent of respondents agreed that a tipping point of 30 percent of corporate audits performed by AI will be achieved by
2025 (World Economic Forum 2015).
While the idea of employing artificial intelligence in accounting and audits is certainly not new (Keenoy 1958), there is
reason to expect that its impact on the field will be more substantial in future years because of recent developments in
information and technology. Artificial intelligence requires both substantial data and processing power, and both are available
in large quantities today. In addition, both open source and proprietary versions of artificial intelligence software have
proliferated over the past several years. Artificial intelligence has gone through several ‘‘winters’’ and ‘‘springs,’’ but this spring
is seeing a larger flowering of activity than ever before.
There are both demand and supply factors behind the latest upsurge in AI. On the demand side, there has not been much
productivity improvement in advanced economies over the past several years (only 1.3 percent average annual growth from
2007 to 2015, and decreasing productivity in the first two quarters of 2016), and companies are anxious to learn whether
cognitive technologies can finally spur productivity growth. There are also many situations today in which a traditional human
approach to analytics and decision-making is simply impossible. These decisions need to be made with too much data and in
too short a time for humans to be employed in the process. Digital advertising, medical diagnosis, predictive maintenance for
industrial equipment, and a detailed audit of all company transactions fall into this category.
On the supply side, we now have both software and hardware that is well suited to performing cognitive tasks. Both
proprietary and open source software is widely available to perform various types of machine cognition. Google, Microsoft,
Facebook, and Yahoo! have all made available open source machine learning libraries. And some of the world’s largest IT
companies are providing proprietary offerings.
Data scientists, however, often comment that the supply-side factors that really make a difference for this generation of AI
are data and processing power. Neural networks, for example, have been available since the 1950s. But current versions of
them—some of which are called ‘‘deep learning’’ because they have multiple layers of features or variables to make a decision

Editor’s note: Accepted by Miklos A. Vasarhelyi.

Submitted: December 2016
Accepted: February 2017
Published Online: April 2017

115

Kokina and Davenport

116

TABLE 1
Aggregate Task Structure
Adapted from Abdolmohammadi (1999)
No. of
Tasks

Audit Phase
Orientation

45

Control Structure

75

Substantive Tests

171

Forming an Opinion and Financial Statement Reporting
Total

41
332

Task Structure
Structured

Semi-Structured

Unstructured

7
(16%)
10
(13%)
114
(67%)
0
(0%)

14
(31%)
58
(77%)
54
(32%)
9
(22%)

24
(53%)
7
(10%)
3
(1%)
32
(78%)

131
(39%)

135
(41%)

66
(20%)

about something—require massive amounts of data to learn on, and massive amounts of computing power to solve the complex
problems they address. In many cases, there are data sources at the ready for training purposes. The ImageNet database, for
example—a research database used for training cognitive technologies to recognize images—has over 14 million images from
which a deep learning system can learn.
The availability of almost unlimited computing capability in the cloud means that researchers, application developers, and
accountants can readily obtain the horsepower they need to crunch data with cognitive tools. And relatively new types of
processors like graphics processing units (GPUs) are particularly well suited to addressing some cognitive problems. GPUs in
the cloud provide virtually unlimited processing power for many cognitive applications.
Auditing is particularly suited for applications of data analytics and artificial intelligence because it has become
challenging to incorporate the vast volumes of structured and unstructured data to gain insight regarding financial and
nonfinancial performance of companies. Also, many audit tasks are structured and repetitive and, therefore, can be automated.
In other words, accounting and auditing have not been left behind in this latest AI spring.
Each of the Big 4 accounting firms has invested heavily in technological innovation. KPMG has partnered with IBM’s
Watson AI to develop AI audit tools (Melendez 2016). PricewaterhouseCoopers (PwC) has developed Halo, an analytics
platform that serves as a pipeline to AI and augmented reality products (M2 Presswire 2016). Deloitte has developed Argus for
AI and Optix for data analytics. Traditionally, accounting firms have relied on recent graduates to fill the entry-level positions
required to perform repetitive administrative tasks. Due to the emergence of AI, one Ernst & Young (EY) source estimates that
the number of new hires each year could fall by half, which would substantially alter the industry’s employment model (Agnew
2016).
LITERATURE REVIEW
Artificial Intelligence and Audit Procedures
The focus of AI capabilities in auditing is on the automation of labor-intensive tasks (Rapoport 2016). These are structured
and repetitive tasks performed throughout the audit. The effects of AI are likely to be the most pronounced in audit tasks that
were once performed manually, but have already been supported by some technology (Agnew 2016).
To identify audit areas that are likely to be impacted by AI to the greatest extent, it is important to decompose the audit into
a series of tasks and identify those with the most structure. Abdolmohammadi (1999) provides empirical evidence of audit task
structure based on 49 audit manager and partner evaluations of 332 audit tasks representing six audit phases and 50 subphases.
He finds that the majority of audit work consists of structured tasks (39 percent, or 131 out of 332 tasks) and semi-structured
tasks (41 percent, or 135 out of 332 tasks), with only a small portion of tasks classified as unstructured (20 percent, or 66 out of
332 tasks) (Table 1). Most structured tasks (67 percent, or 114 of 171 tasks) appear to be in the substantive test audit phase.
Furthermore, Abdolmohammadi (1999) reports that the substantive test phase was judged to have the largest proportion of tasks
suitable for decision-aid development.
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

The Emergence of Artificial Intelligence: How Automation is Changing Auditing

117

An analysis of structured automated tasks that are presented in Abdolmohammadi (1999) reveals that automation assists
with tasks that include verification, recomputation, footing, and vouching. Specific examples of several structured automated
tasks performed during the substantive test audit phase are: the verification of mathematical accuracy of all relevant supporting
schedules (Task No. 18); the footing of cash receipts journal and cash disbursement journal and tracing to general ledger
postings and bank statements (Task No. 10); the recomputation of depreciation (amortization) book and tax basis (Task No.
109); the footing of voucher register, tracing to general ledger, inspection of supervisory review, and approval of
summarization and posting (Task No. 76); the footing of purchases journal and tracing to accounts payable subsidiary ledger,
and inspection of supervisory review and approval of summarization and posting (Task No. 77).
Srinivasan (2016) recently published a high-level model of external audit process activities, and argued that automation of
them could lead to the extinction of human auditors, although it was only speculative. Even though there is no detailed model
of which we are aware that identifies specific audit tasks suitable for AI-enabled technology applications, Baldwin, Brown, and
Trinkle (2006) summarize earlier uses of AI in documenting the use of neural networks in the performance of analytical review
procedures and risk assessment, and the use of genetic algorithms to assist with classification tasks (e.g., collectible debt or a
bad debt). They also document the use of expert systems in materiality assessments, internal control evaluations, and going
concern judgments. To more holistically conceptualize a modern-day AI-enabled audit, Issa, Sun, and Vasarhelyi (2016)
envision it to be similar to an assembly line in which an output of one step turns into an input of the step that follows. They
identify seven distinct audit phases from pre-planning to the audit report and present how AI could transform each of those
phases of the audit process.
Currently, the impact of AI in audits is especially pronounced in the area of data acquisition (data extraction, comparison,
and validation) (Brennan, Baccala, and Flynn 2017). This means that AI-enabled technology can locate relevant information,
extract it from documents, and make it usable for the human auditor, who can devote more time to areas requiring higher-level
judgment. For example, AI enables full automation of such time-consuming tasks as payment transaction testing, including
extraction of any supporting data for further substantive testing (Brennan et al. 2017).
Modern AI tools are increasingly able to scan for keywords and patterns in complex electronic documents to identify and
extract relevant accounting information from various sources, such as sales, contracts, and invoices (Agnew 2016). For
example, AI tools can spot if a company records unusually high sales figures just before the end of a reporting period, or
disburses unusually high payments right after the end of the reporting period (Rapoport 2016). AI tools can also spot anomalies
in the data, such as an unexpected order increase in a particular region, unusually high expense items posted by an individual,
or exceptionally favorable equipment lease terms for a supplier (Brennan et al. 2017). Overall, ‘‘as audits become increasingly
automated, there will be less emphasis on ticking and tying and vouching, and greater emphasis on understanding the overall
picture painted by the data, better understanding inputs and assumptions, and identifying and evaluating trends, patterns, and
outliers’’ (Accounting Today 2016).
DISCUSSION
Artificial Intelligence Capabilities in Accounting and Auditing
Artificial intelligence (also known as cognitive technology or cognitive computing, which we view as synonymous terms
with AI) is a broad category, and not all aspects of it are relevant to accounting. In Exhibit 1, we describe a set of tasks that
cognitive technologies perform, and the level of intelligence they have reached thus far (Davenport and Kirby 2016a).
Most of the task categories are relevant to accounting and auditing. Performing physical tasks is the traditional domain of
robots, but it may have relevance to certain auditing tasks like counting inventory. Certainly, ‘‘analyze numbers’’ is the
dominant task in accounting and auditing. This has traditionally meant algebraic analysis, but accountants and auditors are
increasingly using business intelligence and visual analytics to communicate results (Schneider, Dai, Janvrin, Ajayi, and
Raschke 2015). They are also employing hypothesis-based predictive analytics, as well, to predict the likelihood of financial
events and malfeasance (Tschakert, Kokina, Kozlowski, and Vasarhelyi 2016).
When this type of analytics is done on a repetitive operational level, it qualifies as ‘‘repetitive task automation.’’ Some
accounting firms have begun to do this type of ongoing production work in the context of their auditing ‘‘platforms,’’ although
it is only in the early stages of application.
The next level of intelligence for analyzing numbers is machine learning, which is already being widely used outside of
accounting to automate statistical and mathematical modeling. It is particularly relevant when organizations wish to
dramatically increase the speed, granularity, and productivity of modeling. As we note below, it is just beginning to be used by
large accounting firms to analyze data. In particular, it can be used for identifying anomalies in large datasets, which may be a
basis for further forensic investigation.
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

Kokina and Davenport

118

EXHIBIT 1
Types of Cognitive Technology and Their Intelligence Level
Adapted from Davenport and Kirby (2016a)

In the task category of digesting words and images, the most common accounting and auditing application is analyzing
contracts and other financially relevant documents. The first step in this process is the ‘‘human support’’ task of recognizing
characters and translating documents into digital information. If this is done on a large scale, then it qualifies as repetitive task
automation. But a higher level of intelligence is necessary to understand context within the document and extract relevant
details from it. This ‘‘natural language processing’’ capability is available from external vendors and has been adopted by
accounting and law firms. ‘‘Natural language generation,’’ or the automated creation of meaningful text, is being used for
accounting-oriented tasks such as creating ‘‘Suspicious Activity Reports’’ for anti-money-laundering processes in financial
services (Davenport 2016). It could eventually also be used for the automated generation of other required audit reports. As yet,
there are no current accounting-oriented uses of image recognition of which we are aware, but there might be a future use with
regard to recognizing and counting certain types of inventory. As Cathy Engelbert, CEO of Deloitte LLC, put it in a speculative
comment:
This might sound a little sci-fi to you, but drones could do physical inventory observations. Maybe you wouldn’t have
to send people out to look at that kind of thing. Take it one step further. We could use imaging technology to look at
things like storage tanks and grain silos. We could use it for a variety of things as you look at the industrial internet of
big things. (Cohn 2016)
The task category of ‘‘perform digital tasks’’ typically involves orchestrating an online process, accessing data, and making
changes in entries and records. This activity has distinct relevance to accounting and auditing processes, and has been widely
adopted for many years by accounting firms to improve productivity in the management of audits (Banker, Chang, and Kao
2002). Current auditing ‘‘platforms’’ (Whitehouse 2015) are an industry-specific version of ‘‘business process management’’
technology that moves work through a process and keeps track of key data. The next level of capability for this task, ‘‘robotic
process automation,’’ automates structured tasks and draws from multiple information systems sources (Lacity and Willcocks
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

The Emergence of Artificial Intelligence: How Automation is Changing Auditing

119

2016). This technology would appear to be quite useful for automating structured audit processes, but we are not aware of its
actual adoption by accounting firms for external audits or by companies for internal audits.
Note in Exhibit 1 that there are no cognitive technologies that are yet capable of self-aware intelligence, although that level
of artificial general intelligence is widely predicted to arrive at some point in the future (Bostrom 2014). That degree of reason
and sentience, sometimes called ‘‘the singularity’’ (Kurzweil 2006), is seen often in fictional depictions of machines and robots,
and includes such attributes as formulating goals and objectives, using imagination, having broad general intelligence, and
being critical of others’ and one’s own performance. Such machines would be able of accomplishing not just a single task better
than humans, but a broad range of tasks. Most researchers do not predict the arrival of self-aware intelligence in less than 20
years from now, and some speculate that it will take as long as 100 years to arrive (Goertzel 2007).
It also seems likely that these technologies will converge in the future as they achieve greater levels of intelligence. For
example, the ability to digest text and images may well be incorporated into the performance of digital tasks. As another
example, the power of server-based machine learning and natural language processing (for example, IBM’s Watson) are
already being incorporated into physical robots (Wang 2016). All of these developments would be likely to have dramatic
effects on accounting and auditing processes in that technologies would then be able to perform virtually all tasks that humans
do today.
Industry Examples of Artificial Intelligence Usage
The most important evidence of AI’s relevance to accounting is adoption of the technology by practicing accountants and
auditors. Although it is still early in the process, several leading firms have adopted cognitive technology already. Some are still
in development, while others have applied it to production audit processes. Some firms are employing predictive and other
forms of analytics to, for example, examine and summarize entire populations of auditable entities like inventories, rather than
samples. While this technology is an important precursor of cognitive technology, we do not consider it to be cognitive or AI
unless it is autonomous and learns over time.
There appear to be two key strategies that accounting firms are employing to add AI or cognitive capabilities to their
businesses. One involves adoption of a broad set of AI capabilities from one vendor—specifically, IBM’s Watson. KPMG is
the most public proponent of this approach, signing a broad agreement with IBM in March 2016 that is intended to apply
Watson to a variety of audit processes (Lee 2016). The specific audit processes being addressed by the system are not clear, but
Watson has a wide variety of application program interfaces (APIs) that do everything from document entity extraction to facial
recognition. KPMG is also making use of advanced predictive analytics technology from the auto racing firm McLaren Applied
Technologies (Sinclair 2015), although this does not appear to be a cognitive application. Its primary purpose is to examine
financial statement risk.
The other primary approach is to assemble a variety of cognitive capabilities from diverse vendors and integrate them as
required into an audit process and platform. This is the primary approach taken by Deloitte, for example. According to Jon
Raphael (2016), the firm’s Chief Innovation Officer, Deloitte has focused on several specific audit subprocesses and tasks,
including:







document review
confirmations
inventory counts
disclosure research
predictive risk analytics
client request lists

Some of these activities are more exploratory, and others are already in production. For example, Deloitte partnered with
the vendor Kira Systems to do document review and extract the relevant terms from contracts, leases, employment agreements,
invoices, and other legal documents. The system learns from human interaction and improves its ability to extract important and
relevant information over time (Whitehouse 2015). By March of 2016, the Kira-based solution (which Deloitte calls ‘‘Argus’’)
had been applied to over 100,000 documents (Kepes 2016).
Raphael (2016) also suggests that additional tasks are forthcoming, and that the most challenging aspect of applying AI to
audits is getting the data in a structured and consistent format across clients. Deloitte is exploring the use of machine learning
technology for the integration and structuring of data.
PwC and EY, the other two members of the Big 4 accounting firms, appear to be making increasing use of audit platforms
and predictive analytics, but not the higher levels of intelligence and cognitive capability, as described in Exhibit 1. PwC, for
example, employs ‘‘Halo’’ for analyzing accounting journals. Most of the analysis is traditional human support-based business
intelligence, but there are some automated algorithms, as well (PwC 2016). EY is focused primarily on Big Data and analytics
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

Kokina and Davenport

120

in audits (EY 2016), with a focus on ‘‘delivering audit analytics by processing large client datasets within our environment,
integrating analytics into our audit approach and getting companies comfortable with the future of audit.’’
Likely Implications for Human Accountants
Senior accountants in large firms uniformly argue that the need for human accountants will not go away anytime soon
(Agnew 2016). But many argue that the skills for successful accounting and auditing are likely to be different in the future, and
some admit that they will need substantially fewer entry-level accountants in coming years.
At least over the next couple of decades, accounting is one of the many business fields that is likely to be augmented by
technology rather than fully automated (Davenport and Kirby 2016a). Since AI technologies replace specific tasks rather than
entire jobs, loss of employment in the short term is likely to be relatively slow and to be marginal rather than dramatic.
Remaining jobs in accounting are likely to involve some of the following types of activities:






Working alongside intelligent accounting machines to monitor their performance and results, and (if possible) to
improve their performance;
Overseeing the use of intelligent machines in external and internal audit processes, and determining whether more, less,
or different automation tools are necessary;
Working with accounting firms and vendors to develop new AI-based technologies, and to support existing ones;
Carrying out tasks that are now impossible with AI-based computers, including cultivating internal and external clients,
interpreting audit and financial results for senior managers and boards of directors, and so forth;
Addressing types of accounting tasks that are so narrow and uncommon that it would be uneconomical to build systems
to automate them.

These five types of remaining jobs for human accounts correspond to five augmentation roles described in the augmentation
literature (Davenport and Kirby 2016b).
It has already been noted (Tschakert et al. 2016) that many accounting programs do not currently prepare students for such
roles. In addition, since many of the remaining tasks will require an understanding of the client’s business and the ability to
communicate effectively with clients, job roles that persist will probably be held by those accountants with substantial
experience. Since the tasks performed by entry-level accountants are relatively structured today, they are the most likely ones to
be automated. As in other professions, such as law and architecture, entry-level students will probably bear the brunt of
automation’s impact on the accounting labor market. This also raises the issue of how accountants will accumulate experience
if there are substantially fewer recruits entering firms just out of school.
IMPLICATIONS FOR FUTURE RESEARCH
It is important for future research to examine bias in AI and whether humans using AI applications can engage in
appropriate judgment and decision-making. Hammond (2016) points to the lack of objectivity and cautions that when
intelligent machines are deployed, they tend to reflect the biases of humans who create or interact with them. The first bias is
data-driven bias, which is associated with the systems generating biased outcomes because of the flaws or skewness in the
underlying data. Another bias is bias through interaction that occurs when machines learn the biases of the people who train
them. Emergent bias is the ‘‘algorithmic version of ‘confirmation bias.’’’ It takes place when machines shield humans from
conflicting points of view while providing them with information that confirms their preferences or beliefs (i.e., personalization
bubble). Finally, a conflicting-goals bias is an unforeseen bias that occurs as a result of a stereotype-driven human interaction
with the system.
Another direction for future research could address the role of transparency, or the lack thereof, in AI-based accounting and
auditing decisions. Earlier versions of AI (for example, rule-based expert systems) and analytics (linear regression analysis)
made it relatively easy for human observers to understand the relationships between inputs, transformations, and outputs of
models. However, machine learning and deep learning neural networks, for example, are often ‘‘black boxes’’ that are difficult
or impossible to understand and interpret, even for technical experts. Until such technologies are made more transparent, it may
be difficult for regulatory bodies, accounting firms, and audited organizations to turn over decisions and judgments to them.
Research on the challenges this poses and some possible solutions to it would be helpful to the field.
Future research should examine how these bias and transparency issues are addressed on behalf of both intelligent system
creators and users in the context of accounting and auditing. Will the benefits of AI auditing systems outweigh the unintended
consequences of the potential biases and uninterpretability? To what extent will human auditors rely on the results of tasks
completed by intelligent systems? What benefits and problematic issues will be uncovered as this capability matures?
There is plenty of evidence that the role of AI in accounting and auditing is proceeding rapidly. This development has
major potential implications for both the quality and process of accounting and auditing work. Accounting researchers and
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

The Emergence of Artificial Intelligence: How Automation is Changing Auditing

121

practitioners will need to collaborate closely in the coming years to shed more light on this transition and provide guidance to
firms and regulators.

REFERENCES
Abdolmohammadi, M. J. 1999. A comprehensive taxonomy of audit task structure, professional rank and decision aids for behavioral
research. Behavioral Research in Accounting 11: 51–92.
Accounting Today. 2016. The audit of the future. Accounting Today. Available at: https://www.accountingtoday.com/news/the-audit-ofthe-future
Agnew, H. 2016. Auditing: Pitch battle. Financial Times (May 9). Available at: https://www.ft.com/content/268637f6-15c8-11e6-9d9800386a18e39d
Baldwin, A. A., C. E. Brown, and B. S. Trinkle. 2006. Opportunities for artificial intelligence development in the accounting domain.
Intelligent Systems in Accounting, Finance and Management 14 (3): 77–86. doi:10.1002/isaf.277
Banker, R. D., H. Chang, and Y. Kao. 2002. Impact of information technology on public accounting firm productivity. Journal of
Information Systems 16 (2): 209–222. doi:10.2308/jis.2002.16.2.209
Bostrom, N. 2014. Superintelligence: Paths, Dangers, Strategies. New York, NY: Oxford University Press.
Brennan, B., M. Baccala, and M. Flynn. 2017. Artificial intelligence comes to financial statement audits. CFO.com (February 2).
Available at: http://ww2.cfo.com/auditing/2017/02/artificial-intelligence-audits/
Cohn, M. 2016. Deloitte leverages AI and analytics for audits. Accounting Today (November 14). Available at: http://www.
accountingtoday.com/news/deloitte-leverages-ai-and-analytics-for-audits
Davenport, T. H. 2016. Make way for natural language generation. LinkedIn Pulse (February 2). Available at: https://www.linkedin.com/
pulse/make-way-natural-language-generation-tom-davenport
Davenport, T. H., and J. Kirby. 2016a. Just how smart are smart machines? MIT Sloan Management Review (Spring). Available at: http://
sloanreview.mit.edu/article/just-how-smart-are-smart-machines/
Davenport, T. H., and J. Kirby. 2016b. Only Humans Need Apply: Winners and Losers in the Age of Smart Machines. New York, NY:
Harper Business.
Ernst & Young (EY). 2016. How Big Data and Analytics are Transforming the Audit. Available at: http://www.ey.com/gl/en/services/
assurance/ey-reporting-issue-9-how-big-data-and-analytics-are-transforming-the-audit
Goertzel, B. 2007. Human-level artificial general intelligence and the possibility of a technological singularity: A reaction to Ray
Kurzweil’s The Singularity is Near, and McDermott’s critique of Kurzweil. Artificial Intelligence 171 (18): 1161–1173. doi:10.
1016/j.artint.2007.10.011
Hammond, K. 2016. 5 Unexpected Sources of Bias in Artificial Intelligence. Available at: https://techcrunch.com/2016/12/10/5unexpected-sources-of-bias-in-artificial-intelligence/
Issa, H., T. Sun, and M. Vasarhelyi. 2016. Research ideas for artificial intelligence in auditing: The formalization of audit and workforce
supplementation. Journal of Emerging Technologies in Accounting 13 (2): 1–20. doi:10.2308/jeta-10511
Keenoy, C. L. 1958. The impact of automation on the field of accounting. The Accounting Review 33 (2): 230–236.
Kepes, B. 2016. Big four accounting firms delve into artificial intelligence. Computerworld (March 16). Available at: http://www.
computerworld.com/article/3042536/big-data/big-four-accounting-firms-delve-into-artificial-intelligence.html
Kurzweil, R. 2006. The Singularity is Near: When Humans Transcend Biology. New York, NY: Penguin.
Lacity, M. C., and L. Willcocks. 2016. A new approach to automating services. MIT Sloan Management Review (Fall). Available at:
http://sloanreview.mit.edu/article/a-new-approach-to-automating-services/
Lee, D. 2016. KPMG recruits IBM Watson for cognitive tech audits, insights. Accounting Today (March 8). Available at: http://www.
accountingtoday.com/news/kpmg-recruits-ibm-watson-for-cognitive-tech-audits-insights
M2 Presswire. 2016. PwC Wins ‘‘Audit Innovation of the Year’’ at the Accountant & International Accounting Bulletin Awards. Available
at: https://www.m2.com/m2/web/story.php/20166219039
Melendez, C. 2016. Artificial intelligence gets into auditing, what’s next? Infoworld.com (March 16). Available at: http://www.infoworld.
com/article/3044468/application-development/artificial-intelligence-gets-into-auditing-whats-next.html
PricewaterhouseCoopers (PwC). 2016. Halo for Journals. Available at: http://halo.pwc.com/
Raphael, J. 2016. Email correspondence with author (Davenport). (December 22).
Rapoport, M. 2016. Auditors count on tech for backup. Wall Street Journal (March 8).
Schneider, G. P., J. Dai, D. J. Janvrin, K. Ajayi, and R. L. Raschke. 2015. Infer, predict, and assure: Accounting opportunities in data
analytics. Accounting Horizons 29 (3): 719–742. doi:10.2308/acch-51140
Sinclair, N. 2015. How KPMG is using Formula 1 to transform audit. CA Today (October 27). Available at: https://www.icas.com/catoday-news/kpmg-and-formula-one-big-data
Srinivasan, V. 2016. Will financial auditors become extinct? In The Intelligent Enterprise in the Era of Big Data, Chapter 7, pp. 171–183.
New York, NY: Wiley. doi:10.1002/9781118834725.ch7
Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

122

Kokina and Davenport

Sutton, S. G., M. Holt, and V. Arnold. 2016. ‘‘The reports of my death are greatly exaggerated’’—Artificial intelligence research in
accounting. International Journal of Accounting Information Systems 22: 60–73. doi:10.1016/j.accinf.2016.07.005
Tschakert, N., J. Kokina, S. Kozlowski, and M. Vasarhelyi. 2016. The next frontier in data analytics. Journal of Accountancy (August).
Available at: http://www.journalofaccountancy.com/issues/2016/aug/data-analytics-skills.html
Wang, B. 2016. IBM putting Watson into Softbank Pepper robot. Next Big Future (April). Available at: http://www.nextbigfuture.com/
2016/04/ibm-putting-watson-into-softbank-pepper.html
Whitehouse, T. 2015. The technology transforming your annual audit. Compliance Week (December 1). Available at: https://www.
complianceweek.com/news/news-article/the-technology-transforming-your-annual-audit#.WGQkebYrLPA
World Economic Forum. 2015. Deep Shift: Technology Tipping Points and Societal Impact. Available at: http://www3.weforum.org/
docs/WEF_GAC15_Technological_Tipping_Points_report_2015.pdf

Journal of Emerging Technologies in Accounting
Volume 14, Number 1, 2017

Copyright of Journal of Emerging Technologies in Accounting is the property of American
Accounting Association and its content may not be copied or emailed to multiple sites or
posted to a listserv without the copyright holder's express written permission. However, users
may print, download, or email articles for individual use.


Journal of Research in Business, Economics and Management (JRBEM)
ISSN: 2395-2210

SCITECH
RESEARCH ORGANISATION

Volume 8, Issue 3
March 16, 2017

Journal of Research in Business, Economics and Management
www.scitecresearch.com

Exploring the Impact of Artificial Intelligence on the Accounting
Profession
Cindy Greenman, PhD., CFE
Embry-Riddle Aeronautical University – Prescott, AZ

Abstract
Artificial intelligence is no longer the robots and computers of science fiction from Hollywood movies. The
ideas of developing machines that can “learn” are centuries old. The capacities of the computers and software
of today create and exhibit intelligence, but also bring with it concerns along with much promise. In the
accounting field artificial intelligence has been taking on more and more tasks. Already, there is software that
has automated many accounting, tax, bookkeeping, and auditing processes. If machines are assuming a
greater role, where do we, the professionals strike a balance? What does the future of the accounting
profession look like with the growth of artificial intelligence?

Keywords: Accounting; auditing; artificial intelligence; expert systems; machine learning; knowledge based
systems; cognitive systems.

Introduction
Accountants and auditors are responsible for preparing and examining the financial records of companies. They ensure
that the records are accurate, that taxes are paid in a timely manner and for the proper amount. They also analyze
financial operations and try to help the organizations to run in a more efficient manner. The field of accounting has a long
history of artificial intelligence (AI) applications dating back more than 25 years mainly in the areas of financial
reporting and auditing tasks. According to research done by the University of Oxford in 2015, accountants have a 95
percent change of losing their jobs as machines take over the role of data analytics and number crunching. However, this
same report found that as technology progresses, some jobs are eliminated while others are created. 1

History of AI
The concept of intellectual machines can be traced as far back as Greek mythology. Greek myths contain stories of
Hephaestus, a blacksmith who contrived mechanical robots. Other myths include mechanical toys and human-like
androids. Intelligent relics begin appearing in literature since that time. In the 4th century B.C. Aristotle created the first
formal deductive reasoning system (syllogistic logic). In 1206 A.D. an Arab inventor built what is believed to be the first
programmable humanoid robot. By the 17 th century Pascal was creating the first calculator (1642). The first “computer”
game based on the game of chess came along in the early 20 th century (1912). It was in 1936 Alan Turing first suggested
the idea of the Touring Machine. This machine was the basis for theories about computers and computing. 2
With the development of stored-program computers in the mid-20th century the realistic concept of artificial intelligence
really begins. In 1956 at the first conference devoted to the subject of “artificial intelligence”. In 1961 UNIMATE, the
first mass-produced industrial robot started working at General Motors beginning the revolution of factory automation.
This robot did the work that was deemed harmful to humans. By 1969 GM was producing 110 cars per hour, which was
more than double the rate of any other automotive facility in existence at that time.
1
2

Griffin, O. (October, 2016). How artificial intelligence will impact accounting. Economia.
AI Topics. (2016). Brief History of Artificial Intelligence. http://aitopics.org/topic/brief-histories-timelines.

Volume 8, Issue 3 available at www.scitecresearch.com/journals/index.php/jrbem/index

1451

Many blame the automation revolution for the reduction in manufacturing jobs. Some say that the growth of production
came at the cost of manufacturing jobs. Graetz and Michaels did a study in 2014 that showed the United States had
increased the use of automation per hour worked by 237 percent between 1993 and 2007. During the same time frame
there were 2.2 million manufacturing jobs lost. Their study showed that there was essentially no relationship between the
number of jobs lost and the amount of automation. It showed that if there was a correlation between the two the U.S.
should have lost an additional one-third than was really lost.3

AI Technology in Accounting
The job description of today’s accountant looks very different than that of the accountant of 20 years ago. In another 20
years, accountants will again, play a different role. Their roles will change substantially over the next decade. More
emphasis will be placed on consulting, business development, advisory services and risk management. Accountants will
need to embrace specialization and the use of technology. 4
Artificial intelligence is being designed to think, feel, and react like a living, breathing creature. According to a study
done by Deloitte, AI could emerge with a whole new class of products and services specifically applicable in the areas of
accounting. These include: customer service, research and development, logistics, sales, marketing and informational
analysis. For those professions that require following specific methodologies, information analysis, report preparation,
and many cumbersome processes (i.e. bookkeeping, transaction coding, etc.), AI has the potential to completely alter the
profession. According to a study done by the Association of Chartered Certified Accountants, there is the possibility that
automation will relieve many burdensome tasks that would enable accountants to focus on consulting services and other
higher-value work. 5
In the very near future, AI may be completely involved in the monitoring and evaluating of compliance with regulations,
organizational policy, employee evaluations and even hiring and firing. Obviously, accounting software is not new to the
profession. Tax filing software has not put accountants out of business, it has, in fact, made them more efficient and
made it possible to file many more returns than they could before. However, the new incoming software could likely
empower some users to the point when they will not need their accountant any longer. The latest evolution of products
are more “cloud” based, such as the QuickBooks Online, which seems to compel some to take on some of the
bookkeeping tasks of their business.
There are differing opinions on how the role of an accountant will change. Some are of the opinion that there will be a
major modification as was the case in the taxi/transportation industry when Uber and Lyft were introduced. Others
believe that software will simply shift some of the less complicated tasks to the businesses themselves, but that they will
still be in need of credentialed experts to conduct audits and sort through the highly complex regulations. 6

AI in Auditing
Cognitive technologies actually further the power of information technology to those tasks that traditionally are
performed by humans, they enable users to shatter what was once a tradeoff between speed, cost, and quality. These AI
technologies can facilitate auditors to automate those tasks that have been conducted manually by humans for decades.
As a result, auditors can be freed in order to focus on improving quality by evaluating advanced analytics, spending
additional time providing insight and applying better professional judgement. One particular area that AI has been
extremely useful is that of document review. Reading through pages and pages of contracts in order to mine key terms
has, in the past, been a time intensive, manual process. Using artificial intelligence this concept has now become
automated. The “learning technology” that is possible with this type of processing is making it feasible to train the
system on a set of sample documents so that the system then learns how to identify and extract key terms. 7
In 20116 KPMG released plans to begin using artificial intelligence on their audit engagements in Australia. Their
proposal is to use IBM’s cognitive computing technology called “Watson”. Executives from KPMG maintain that by
using Watson they can extend the data and analytics. Where sample sizes were once limited by time and man-power,
there is now no limitation to the sampling that can be done. Rather than simply analyzing a sampling of total data,
KPMG will be able to scrutinize all of the numbers. More data being analyzed means better comprehension for the
3

Graetz, G and Michaels, G. (2015). Estimating the impact of robots on productivity and employment. Center for
Economic Performance. http://cep.lse.ac.uk/pubs/download/dp1335.pdf.
4
McCabe, S. (2014). CPA.com study gauges firms’ preparedness for the future. Accounting Tomorrow.
http://www.accountingtoday.com/blogs/accounting-tomorrow/cpa-com-study-gauges-firms-readiness-for-the-future73011-1.html
5
Jariwala, B. (2015). Exploring Artificial Intelligence & the Accountancy Profession: Opportunity, Threat, Both, Neither?
International Federation of Accountants.
6
Poston, J. (2014). Can Software Really Replace Accountants? AccountingWeb.com
7
Raphael, J. (2015). How Artificial Intelligence Can Boost Audit Quality. CFO.com

Volume 8, Issue 3 available at www.scitecresearch.com/journals/index.php/jrbem/index

1452

clients and (more than likely) bigger audit fees. KPMG, which ranks fourth among accounting firms in almost all areas
(revenue, number of employees, salaries and assets). Of course KPMG is not alone in using this type of technology.
Deloitte, Ernst & Young and PriceWaterhouseCooper (PwC) are employing similar technologies, only on a smaller
scale.8

Job Growth (Number of Accountants)
It is difficult to find an exact number of accountants in the United States because many are not registered. However, in
2016 there were 664,532 CPAs (Certified Public Accountants) in the US.9 According to the United States Department of
Labor, Bureau of Labor Statistics there are 1,226,910 individuals employed in the accounting profession in the United
States. Many fear that with the advances in technology, specifically in artificial intelligence, there will be a loss of jobs.
The fear is that computers will take over for humans, offering free labor, better accuracy and no personality conflicts. If
those fears were being realized we’d expect to see a decline in the number of professional accountants. However, the
exact opposite is true. The Bureau of Labor and Statistics reports that the accounting profession is projected to grow at a
rate of 11 percent over the next 10 years, an increase of over 142,000 new accounting and auditing jobs. 10 Some of this
may be due to the openings left by the Boomer generation retiring, but not all of it can be attributed to that one factor.
According to the Accounting and Financial Women’s Alliance, businesses are modifying their business models to take
advantage of big data and taking a more analytical view. Chief Financial Officers (CFOs) are looking to hire finance and
accounting individuals who are experienced in data analytics, modeling techniques, proficient with accounting software
and advanced in Microsoft Excel. In other words, they are looking for people who can work with the new technology.
They want people with the skills necessary to work in a global company and that can keep up with the quickly changing
demands of technology. 11

Conclusion
Artificial Intelligence is critical to the future of the accounting and auditing professions. AI is a vital tool that will
provide these professionals with the needed tools to increase the efficiency and effectiveness of their occupations.12 The
repetitive tasks of bookkeeping or process-driven assignments are more likely to be replaced with an automated
technology than the higher value specialties that involve professional judgement. Many believe that the younger
generation of accountants need to understand and be prepared to work alongside artificial intelligence.
So, are we finally at the point of machines taking over our world? Online education taking over for professors, investing
websites taking over for personal financial advisors, legal software taking over for lawyers, the list goes on. The
accounting profession is not immune to this phenomenon of new technologies disrupting the workforce. The use of tax
filing software hasn’t put accountants out of business, it simply changed the number of tax returns an accountant was
able to prepare. Quick Books has not reduced the income of accountants, it simply changed the focus from paper and
pencil entry, to computer and software entry. AI in the accounting world will not replace accountants, it will simply
change the focus.

References

8
9

[1]

Griffin, O. (October, 2016). How artificial intelligence will impact accounting. Economia.

[2]

AI Topics. (2016). Brief History of Artificial Intelligence. http://aitopics.org/topic/brief-histories-timelines.

[3]

Graetz, G and Michaels, G. (2015). Estimating the impact of robots on productivity and employment. Center for
Economic Performance. http://cep.lse.ac.uk/pubs/download/dp1335.pdf.

[4]

McCabe, S. (2014). CPA.com study gauges firms’ preparedness for the future. Accounting Tomorrow.
http://www.accountingtoday.com/blogs/accounting-tomorrow/cpa-com-study-gauges-firms-readiness-for-thefuture-73011-1.html

[5]

Jariwala, B. (2015). Exploring Artificial Intelligence & the Accountancy Profession: Opportunity, Threat, Both,
Neither? International Federation of Accountants.

Pash, C. (2016). KPMG will soon be using artificial intelligence for audits in Australia. Business Insider.
Membership (2016). American Institute of Certified Public Accountants.

10

Best Jobs. (2016). US News and World Report. http://money.usnews.com/careers/bestjobs/rankings/bestbusiness-jobs.
11

A Salary Guide Foresees a Cheery Future for Accounting and Finance Professionals. (2014). Accounting & Financial
Women’s Alliance.
12
Baldwin, A., Brown, C., & Trinkle, B. (2006). Opportunities for Artificial Intelligence Development in the Accounting
Domain: The Case for Auditing.

Volume 8, Issue 3 available at www.scitecresearch.com/journals/index.php/jrbem/index

1453

[6]

Poston, J. (2014). Can Software Really Replace Accountants?

[7]

Raphael, J. (2015). How Artificial Intelligence Can Boost Audit Quality. CFO.com

[8]

Pash, C. (2016). KPMG will soon be using artificial intelligence for audits in Australia. Business Insider.

[9]

Membership (2016). American Institute of Certified Public Accountants.

[10] Best
Jobs.
(2016).
US
jobs/rankings/bestbusiness-jobs.

News

and

World

Report.

http://money.usnews.com/careers/best-

[11] A Salary Guide Foresees a Cheery Future for Accounting and Finance Professionals. (2014). Accounting &
Financial Women’s Alliance.
[12] Baldwin, A., Brown, C., & Trinkle, B. (2006). Opportunities for Artificial Intelligence Development in the
Accounting Domain: The Case for Auditing.

Volume 8, Issue 3 available at www.scitecresearch.com/journals/index.php/jrbem/index

1454


A Berkeley View of Systems Challenges for AI
Ion Stoica, Dawn Song, Raluca Ada Popa, David Patterson, Michael W. Mahoney, Randy Katz,
Anthony D. Joseph, Michael Jordan, Joseph M. Hellerstein, Joseph Gonzalez, Ken Goldberg,
Ali Ghodsi, David Culler, Pieter Abbeel∗

arXiv:1712.05855v1 [cs.AI] 15 Dec 2017

ABSTRACT
With the increasing commoditization of computer vision, speech
recognition and machine translation systems and the widespread
deployment of learning-based back-end technologies such as digital advertising and intelligent infrastructures, AI (Artificial Intelligence) has moved from research labs to production. These
changes have been made possible by unprecedented levels of data
and computation, by methodological advances in machine learning,
by innovations in systems software and architectures, and by the
broad accessibility of these technologies.
The next generation of AI systems promises to accelerate these
developments and increasingly impact our lives via frequent interactions and making (often mission-critical) decisions on our behalf,
often in highly personalized contexts. Realizing this promise, however, raises daunting challenges. In particular, we need AI systems
that make timely and safe decisions in unpredictable environments,
that are robust against sophisticated adversaries, and that can process ever increasing amounts of data across organizations and individuals without compromising confidentiality. These challenges
will be exacerbated by the end of the Moore’s Law, which will constrain the amount of data these technologies can store and process.
In this paper, we propose several open research directions in systems, architectures, and security that can address these challenges
and help unlock AI’s potential to improve lives and society.

KEYWORDS
AI, Machine Learning, Systems, Security

1

foster new industries around IoT, augmented reality, biotechnology
and autonomous vehicles.
These applications will require AI systems to interact with the
real world by making automatic decisions. Examples include autonomous drones, robotic surgery, medical diagnosis and treatment,
virtual assistants, and many more. As the real world is continually changing, sometimes unexpectedly, these applications need to
support continual or life-long learning [96, 109] and never-ending
learning [76]. Life-long learning systems aim at solving multiple
tasks sequentially by efficiently transferring and utilizing knowledge from already learned tasks to new tasks while minimizing
the effect of catastrophic forgetting [71]. Never-ending learning is
concerned with mastering a set of tasks in each iteration, where
the set keeps growing and the performance on all the tasks in the
set keeps improving from iteration to iteration.
Meeting these requirements raises daunting challenges, such
as active exploration in dynamic environments, secure and robust
decision-making in the presence of adversaries or noisy and unforeseen inputs, the ability to explain decisions, and new modular
architectures that simplify building such applications. Furthermore,
as Moore’s Law is ending, one can no longer count on the rapid
increase of computation and storage to solve the problems of nextgeneration AI systems.
Solving these challenges will require synergistic innovations
in architecture, software, and algorithms. Rather than addressing
specific AI algorithms and techniques, this paper examines the
essential role that systems will play in addressing challenges in AI
and proposes several promising research directions on that frontier.

INTRODUCTION

Conceived in the early 1960’s with the vision of emulating human
intelligence, AI has evolved towards a broadly applicable engineering discipline in which algorithms and data are brought together
to solve a variety of pattern recognition, learning, and decisionmaking problems. Increasingly, AI intersects with other engineering and scientific fields and cuts across many disciplines in computing.
In particular, computer systems have already proved essential in
catalyzing recent progress in AI. Advances in parallel hardware [31,
58, 90] and scalable software systems [32, 46, 114] have sparked the
development of new machine learning frameworks [14, 31, 98] and
algorithms [18, 56, 62, 91] to allow AI to address large-scale, realworld problems. Rapidly decreasing storage costs [1, 80], crowdsourcing, mobile applications, internet of things (IoT), and the competitive advantage of data [40] have driven further investment in
data-processing systems and AI technologies [87]. The overall effect
is that AI-based solutions are beginning to approach or even surpass
human-level capabilities in a range of real-world tasks. Maturing AI
technologies are not only powering existing industries—including
web search, high-speed trading and commerce—but are helping to

2

WHAT IS BEHIND AI’S RECENT SUCCESS

The remarkable progress in AI has been made possible by a “perfect
storm” emerging over the past two decades, bringing together:
(1) massive amounts of data, (2) scalable computer and software
systems, and (3) the broad accessibility of these technologies. These
trends have allowed core AI algorithms and architectures, such as
deep learning, reinforcement learning, and Bayesian inference to
be explored in problem domains of unprecedented scale and scope.

2.1

Big data

With the widespread adoption of online global services, mobile
smartphones, and GPS by the end of 1990s, internet companies
such as Google, Amazon, Microsoft, and Yahoo! began to amass
huge amounts of data in the form of audio, video, text, and user
logs. When combined with machine learning algorithms, these
massive data sets led to qualitatively better results in a wide range of
core services, including classical problems in information retrieval,
information extraction, and advertising [49].

A Berkeley View of Systems Challenges for AI,

2.2

Big systems

Processing this deluge of data spurred rapid innovations in computer and software systems. To store massive amounts of data, internet service companies began to build massive-scale datacenters,
some of which host nearly 100, 000 servers, and provide EB [65] of
storage. To process this data, companies built new large-scale software systems able to run on clusters of cheap commodity servers.
Google developed MapReduce [32] and Google File System [43], followed shortly by the open-source counterpart, Apache Hadoop [7].
Then came a plethora of systems [46, 55, 60, 67, 114], that aimed to
improve speed, scale, and ease of use. These hardware and software
innovations led to the datacenter becoming the new computer [11].
With the growing demand for machine learning (ML), researchers and practitioners built libraries on top of these systems
to satisfy this demand [8, 52, 75].
The recent successes of deep learning (DL) have spurred a new
wave of specialized software systems have emerged to scale out
these workloads on CPU clusters and take advantage of specialized hardware, such as GPUs and TPUs. Examples include TensorFlow [2], Caffe [57], Chainer [20], PyTorch [89], and MXNet [22].

2.3

Accessibility to state-of-the-art technology

The vast majority of systems that process data and support AI workloads are built as open-source software, including Spark [114], TensorFlow [2], MXNet [22], Caffe [57], PyTorch [89], and BigDL [15].
Open source allows organizations and individuals alike to leverage
state-of-the-art software technology without incurring the prohibitive costs of development from scratch or licensing fees.
The wide availability of public cloud services (e.g., AWS, Google
Cloud, and MS Azure) allows everyone to access virtually unlimited amounts of processing and storage without needing to build
large datacenters. Now, researchers can test their algorithms at a
moment’s notice on numerous GPUs or FPGAs by spending just a
few thousands of dollars, which was unthinkable a decade ago.

3

TRENDS AND CHALLENGES

While AI has already begun to transform many application domains,
looking forward, we expect that AI will power a much wider range
of services, from health care to transportation, manufacturing to
defense, entertainment to energy, and agriculture to retail. Moreover, while large-scale systems and ML frameworks have already
played a pivotal role in the recent success of AI, looking forward,
we expect that, together with security and hardware architectures,
systems will play an even more important role in enabling the broad
adoption of AI. To realize this promise, however, we need to address
significant challenges that are driven by the following trends.

3.1

Mission-critical AI

With ongoing advances in AI in applications, from banking to
autonomous driving to robot-assisted surgery and to home automation, AI is poised to drive more and more mission-critical
applications where human well-being and lives are at stake.
As AI will increasingly be deployed in dynamic environments,
AI systems will need to continually adapt and learn new “skills”
as the environment changes. For example, a self-driving car could
quickly adapt to unexpected and dangerous road conditions (e.g.,

an accident or oil on the road), by learning in real time from other
cars that have successfully dealt with these conditions. Similarly,
an AI-powered intrusion-detection system must quickly identify
and learn new attack patterns as they happen. In addition, such
mission-critical applications must handle noisy inputs and defend
against malicious actors.
Challenges: Design AI systems that learn continually by interacting with a dynamic environment, while making decisions that are
timely, robust, and secure.

3.2

Personalized AI

From virtual assistants to self-driving cars and political campaigns,
user-specific decisions that take into account user behavior (e.g., a
virtual assistant learning a user’s accent) and preferences (e.g., a
self-driving system learning the level of “aggressiveness” a user is
comfortable with) are increasingly the focus. While such personalized systems and services provide new functionality and significant
economic benefits, they require collecting vast quantities of sensitive personal information and their misuse could affect users’
economic and psychological wellbeing.
Challenges: Design AI systems that enable personalized applications and services yet do not compromise users’ privacy and security.

3.3

AI across organizations

Companies are increasingly leveraging third-party data to augment
their AI-powered services [27]. Examples include hospitals sharing data to prevent epidemic outbreaks and financial institutions
sharing data to improve their fraud-detection capabilities. The proliferation of such applications will lead to a transition from data
silos—where one company collects data, processes it, and provides
the service—to data ecosystems, where applications learns and
make decisions using data owned by different organizations.
Challenges: Design AI systems that can train on datasets owned
by different organizations without compromising their confidentiality,
and in the process provide AI capabilities that span the boundaries of
potentially competing organization.

3.4

AI demands outpacing the Moore’s Law

The ability to process and store huge amounts of data has been one
of the key enablers of the AI’s recent successes (see Section 2.1).
However, keeping up with the data being generated will become
increasingly difficult due to the following two trends.
First, data continues to grow exponentially. A 2015 Cisco white
paper [25] claims that the amount of data generated by Internet of
Everything (IoE) devices by 2018 to be 400ZB, which is almost 50x
the estimated traffic in 2015. According to a recent study [100], by
2025, we will need a three-to-four orders of magnitude improvement in compute throughput to process the aggregate output of all
genome sequencers in the world. This would require computation
resources to at least double every year.
Second, this explosion of data is coming at a time when our
historically rapidly improving hardware technology is coming to a
grinding halt [53]. The capacity of DRAMs and disks are expected
to double just once in the next decade, and it will take two decades
before the performance of CPUs doubles. This slowdown means that
storing and processing all generated data will become impracticable.

A Berkeley View of Systems Challenges for AI
Challenges: Develop domain-specific architectures and software
systems to address the performance needs of future AI applications
in the post-Moore’s Law era, including custom chips for AI workloads, edge-cloud systems to efficiently process data at the edge, and
techniques for abstracting and sampling data.

4

RESEARCH OPPORTUNITIES

This section discusses the previous challenges from the systems
perspective. In particular, we discuss how innovations in systems,
security, and architectures can help address these challenges. We
present nine research opportunities (from R1 to R9), organized
into three topics: acting in dynamic environments, secure AI, and
AI-specific architectures. Figure 1 shows the most common relationships between trends, on one hand, and challenges and research
topics, on the other hand.
Trends
Mission-critical AI

R 1, R

2, R 3

5
,R
R4
9
,R
R7

Personalized AI

Challenges & Research

R4,

,R

R7

R5,

R6

8

AI across
organizations

AI demands outpacing
Moore’s Law

R6
R8
,R
R7, R8,

9
R9

Acting in dynamic
environments:
R1: Continual learning
R2: Robust decisions
R3: Explainable decisions
Secure AI:
R4: Secure enclaves
R5: Adversarial learning
R6: Shared learning on
confidential data
AI-specific architectures:
R7: Domain specific hardware
R8: Composable AI systems
R9: Cloud-edge systems

Figure 1: A mapping from trends to challenges and research topics.

4.1

Acting in dynamic environments

Many future AI applications will operate in dynamic environments,
i.e., environments that may change, often rapidly and unexpectedly,
and often in non-reproducible ways. For example, consider a group
of robots providing security for an office building. When one robot
breaks or a new one is added, the other robots must update their
strategies for navigation, planning, and control in a coordinated
manner. Similarly, when the environment changes, either due to the
robots’ own actions or to external conditions (e.g., an elevator going
out of service, or a malicious intruder), all robots must re-calibrate
their actions in light of the change. Handling such environments
will require AI systems that can react quickly and safely even to
scenarios that have not been encountered before.
R1: Continual learning. Most of today’s AI systems, including
movie recommendation, image recognition, and language translation, perform training offline and then make predictions online.
That is, the learning performed by the system does not happen
continually with the generation of the data, but instead it happens
sporadicallly, on very different and much slower time scales. Typically, models are updated daily, or in the best case hourly, while
predictions/decisions happen at second or sub-second granularity.
This makes them a poor fit for environments that change continually and unexpectedly, especially in mission-critical applications.

A Berkeley View of Systems Challenges for AI,
These more challenging environments require agents that continually learn and adapt to asynchronous changes.
Some aspects of learning in dynamic environments are addressed
by online learning [17], in which data arrive temporally and updates
to the model can occur as new data arrive. However, traditional
online learning does not aim to handle control problems, in which
an agent’s actions change the environment (e.g., as arise naturally
in robotics), nor does it aim to handle cases in which the outcomes
of decisions are delayed (e.g., a move in a game of chess whose
outcome is only evaluated at the end, when the game is lost or won).
These more general situations can be addressed in the framework of Reinforcement Learning (RL). The central task of RL is
to learn a function—a “policy”—that maps observations (e.g., car’s
camera inputs or user’s requested content) to actions (e.g., slowing
down the car or presenting an ad) in a sequence that maximizes
long-term reward (e.g., avoiding collisions or increasing sales). RL
algorithms update the policy by taking into account the impact of
agent’s actions on the environment, even when delayed. If environmental changes lead to reward changes, RL updates the policy
accordingly. RL has a long-standing tradition, with classical success
stories including learning to play backgammon at level of the best
human players [108], learning to walk [105], and learning basic
motor skills [86]. However, these early efforts require significant
tuning for each application. Recent efforts are combining deep
neural networks with RL (Deep RL) to develop more robust training algorithms that can work for a variety of environments (e.g.,
many Atari games [77]), or even across different application domains, as in the control of (simulated) robots [92] and the learning
of robotic manipulation skills [66]. Noteworthy recent results also
include Google’s AlphaGo beating the Go world champion [95],
and new applications in medical diagnosis [104] and resource management [33].
However, despite these successes, RL has yet to see widescale
real-world application. There are many reasons for this, one of
which is that large-scale systems have not been built with these use
cases in mind. We believe that the combination of ongoing advances
in RL algorithms, when coupled with innovations in systems design,
will catalyze rapid progress in RL and drive new RL applications.
Systems for RL. Many existing RL applications, such as gameplaying, rely heavily on simulations, often requiring millions or
even billions of simulations to explore the solution space and “solve”
complex tasks. Examples include playing different variants of a
game or experimenting with different control strategies in a robot
simulator. These simulations can take as little as a few milliseconds,
and their durations can be highly variable (e.g., it might take a
few moves to lose a game vs. hundreds of moves to win one). Finally, real-world deployments of RL systems need to process inputs
from a variety of sensors that observe the environment’s state, and
this must be accomplished under stringent time constraints. Thus,
we need systems that can handle arbitrary dynamic task graphs,
where tasks are heterogeneous in time, computation, and resource
demands. Given the short duration of the simulations, to fully utilize a large cluster, we need to execute millions of simulations per
second. None of the existing systems satisfies these requirements.
Data parallel systems [55, 79, 114] handle orders of magnitude fewer
tasks per sec, while HPC and distributed DL systems [2, 23, 82]

A Berkeley View of Systems Challenges for AI,
have limited support for heterogeneous and dynamic task graphs.
Hence, we need new systems to support effectively RL applications.
Simulated reality (SR). The ability to interact with the environment is fundamental to RL’s success. Unfortunately, in real-world
applications, direct interaction can be slow (e.g., on the order of seconds) and/or hazardous (e.g., risking irreversible physical damage),
both of which conflict with the need for having millions of interactions before a reasonable policy is learned. While algorithmic
approaches have been proposed to reduce the number of real-world
interactions needed to learn policies [99, 111, 112], more generally
there is a need for Simulated Reality (SR) architectures, in which an
agent can continually simulate and predict the outcome of the next
action before actually taking it [101].
SR enables an agent to learn not only much faster but also much
more safely. Consider a robot cleaning an environment that encounters an object it has not seen before, e.g., a new cellphone. The robot
could physically experiment with the cellphone to determine how
to grasp it, but this may require a long time and might damage the
phone. In contrast, the robot could scan the 3D shape of the phone
into a simulator, perform a few physical experiments to determine
rigidity, texture, and weight distribution, and then use SR to learn
how to successfully grasp it without damage.
Importantly, SR is quite different from virtual reality (VR);
while VR focuses on simulating a hypothetical environment (e.g.,
Minecraft), sometimes incorporating past snapshots of the real
world (e.g., Flight Simulator), SR focuses on continually simulating
the physical world with which the agent is interacting. SR is
also different from augmented reality (AR), which is primarily
concerned with overlaying virtual objects onto real world images.
Arguably the biggest systems challenges associated with SR are
to infer continually the simulator parameters in a changing realworld environment and at the same time to run many simulations
before taking a single real-time action. As the learning algorithm
interacts with the world, it gains more knowledge which can be
used to improve the simulation. Meanwhile, many potential simulations would need to be run between the agent’s actions, using
both different potential plans and making different “what-if” assumptions about the world. Thus, the simulation is required to run
much faster than real time.
Research: (1) Build systems for RL that fully exploit parallelism,
while allowing dynamic task graphs, providing millisecond-level latencies, and running on heterogeneous hardware under stringent deadlines. (2) Build systems that can faithfully simulate the real-world
environment, as the environment changes continually and unexpectedly, and run faster than real time.
R2: Robust decisions. As AI applications are increasingly
making decisions on behalf of humans, notably in mission-critical
applications, an important criterion is that they need to be robust to
uncertainty and errors in inputs and feedback. While noise-resilient
and robust learning is a core topic in statistics and machine learning,
adding system support can significantly improve classical methods.
In particular, by building systems that track data provenance, we
can diminish uncertainty regarding the mapping of data sources to
observations, as well as their impact on states and rewards. We can
also track and leverage contextual information that informs the design of source-specific noise models (e.g., occluded cameras). These
capabilities require support for provenance and noise modeling in

data storage systems. While some of these challenges apply more
generally, two notions of robustness that are particularly important
in the context of AI systems and that present particular systems
challenges are: (1) robust learning in the presence of noisy and adversarial feedback, and (2) robust decision-making in the presence
of unforeseen and adversarial inputs.
Increasingly, learning systems leverage data collected from unreliable sources, possibly with inaccurate labels, and in some cases
with deliberately inaccurate labels. For example, the Microsoft Tay
chatbot relied heavily on human interaction to develop rich natural
dialogue capabilities. However, when exposed to Twitter messages,
Tay quickly took on a dark personality [16].
In addition to dealing with noisy feedback, another research
challenge is handling inputs for which the system was never trained.
In particular, one often wishes to detect whether a query input is
drawn from a substantially different distribution than the training
data, and then take safe actions in those cases. An example of a safe
action in a self-driving car may be to slow down and stop. More
generally, if there is a human in the loop, a decision system could
relinquish control to a human operator. Explicitly training models
to decline to make predictions for which they are not confident,
or to adopt a default safe course of actions, and building systems
that chain such models together can both reduce computational
overhead and deliver more accurate and reliable predictions.
Research: (1) Build fine grained provenance support into AI systems to connect outcome changes (e.g., reward or state) to the data
sources that caused these changes, and automatically learn causal,
source-specific noise models. (2) Design API and language support for
developing systems that maintain confidence intervals for decisionmaking, and in particular can flag unforeseen inputs.
R3: Explainable decisions. In addition to making black-box
predictions and decisions, AI systems will often need to provide
explanations for their decisions that are meaningful to humans.
This is especially important for applications in which there are
substantial regulatory requirements as well as in applications such
as security and healthcare where legal issues arise [24]. Here, explainable should be distinguished from interpretable, which is often
also of interest. Typically, the latter means that the output of the
AI algorithm is understandable to a subject matter expert in terms
of concepts from the domain from which the data are drawn [69],
while the former means that one can identify the properties of the
input to the AI algorithm that are responsible for the particular
output, and can answer counterfactual or “what-if” questions. For
example, one may wish to know what features of a particular organ in an X-ray (e.g., size, color, position, form) led to a particular
diagnosis and how the diagnosis would change under minor perturbations of those features. Relatedly, one may wish to explore what
other mechanisms could have led to the same outcomes, and the
relative plausibility of those outcomes. Often this will require not
merely providing an explanation for a decision, but also considering
other data that could be brought to bear. Here we are in the domain
of causal inference, a field which will be essential in many future AI
applications, and one which has natural connections to diagnostics
and provenance ideas in databases.
Indeed, one ingredient for supporting explainable decisions is the
ability to record and faithfully replay the computations that led to a
particular decision. Such systems hold the potential to help improve

A Berkeley View of Systems Challenges for AI
decision explainability by replaying a prediction task against past
inputs—or randomly or adversarially perturbed versions of past
inputs, or more general counterfactual scenarios—to identify what
features of the input have caused a particular decision. For example,
to identify the cause of a false alarm in a video-based security
system, one might introduce perturbations in the input video that
attenuate the alarm signal (e.g., by masking regions of the image) or
search for closely related historical data (e.g., by identifying related
inputs) that led to similar decisions. Such systems could also lead
to improved statistical diagnostics and improved training/testing
for new models; e.g., by designing models that are (or are not)
amenable to explainability.
Research: Build AI systems that can support interactive diagnostic
analysis, that faithfully replay past executions, and that can help to
determine the features of the input that are responsible for a particular
decision, possibly by replaying the decision task against past perturbed
inputs. More generally, provide systems support for causal inference.

4.2

Secure AI

Security is a large topic, many aspects of which will be central to
AI applications going forward. For example, mission-critical AI
applications, personalized learning, and learning across multiple
organizations all require systems with strong security properties.
While there is a wide range of secur...