Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Understanding AI Technology

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Understanding AI Technology

A concise, practical, and readable overview of Artificial Intelligence


and Machine Learning technology designed for non-technical
managers, officers, and executives
April 2020

By: Greg Allen, Chief of Strategy and Communications


Joint Artificial Intelligence Center (JAIC)
Department of Defense

Foreword by JAIC Director Lt Gen Jack Shanahan

Acknowledgments
The author would like to thank the following individuals for their assistance reviewing earlier drafts
of this document:

• Dr. Jeff Alstott (IARPA)


• Dr. Nate Bastian (Major, U.S. Army, DoD Joint AI Center)
• Dr. Steven L. Brunton (University of Washington)
• Dr. Matthew Daniels (Georgetown University)
• Dr. Ed Felten (Princeton University)
• Mr. Rob Jasper (Pacific Northwest National Laboratory)
• Dr. John Launchbury (Galois, and formerly DARPA)

Disclaimer
The views expressed in this document are those of the author alone and do not necessarily reflect
the position of the Department of Defense or the United States Government.

Website: https://www.ai.mil/
Twitter: @DoDJAIC
LinkedIn: https://www.linkedin.com/company/dod-joint-artificial-intelligence-center/
Understanding Artificial Intelligence Technology 2

FOREWORD BY LIEUTENANT GENERAL JACK SHANAHAN

It is hard for me to describe the steep slope of the learning curve I faced when I
started the Project Maven journey over three years ago. While in many ways I still
consider myself an Artificial Intelligence neophyte today, what I knew about the
subject back then could barely fill the first few lines of a single page in my trusty
notebook. My journey of discovery since then has been challenging, to say the
least. I only wish Greg Allen's guide to "Understanding AI Technology" had been
available to me in late 2016 as we embarked on our first AI/ML pilot project for ISR
full-motion video analysis.

Greg has performed an inestimable service by writing this guide. AI is changing


national security, and it's essential that DoD leaders have a firm grasp of the
technology's building blocks. As I learned back in 2017 and am reminded daily in
my role as the Director of the Joint AI Center (JAIC), AI is not an elixir. It is an
enabler – one that is critical to our future national security. It is important for all of
us to share the same fundamental understanding of AI technology. Greg's guide
balances breadth and depth in just the right way. It is clear, concise, and cogent.
I am confident it will be a valuable resource for everyone in DoD and beyond.

Lieutenant General John N.T. “Jack” Shanahan

Director, Joint Artificial Intelligence Center


Department of Defense
April 2020

Gregory C. Allen | DoD Joint AI Center 2


Understanding Artificial Intelligence Technology 3

EXECUTIVE SUMMARY
Many officials throughout the Department of Defense are asked to make
decisions about AI before they have an appropriate understanding of the
technology’s basics. This guide will help.
The DoD AI Strategy defines AI as “the ability of machines to perform tasks that
normally require human intelligence.” This definition includes decades-old DoD
AI, such as aircraft autopilots, missile guidance, and signal processing systems.
Though many AI technologies are old, there have been legitimate technological
breakthroughs over the past ten years that have greatly increased the diversity of
applications where AI is practical, powerful, and useful. Most of the breakthroughs
and excitement about AI in the past decade have focused on Machine Learning
(ML), which is a subfield of AI. Machine Learning is closely related to statistics and
allows machines to learn from data.
The best way to understand Machine Learning AI is to contrast it with an older
approach to AI, Handcrafted Knowledge Systems. Handcrafted Knowledge
Systems are AI that use traditional, rules-based software to codify subject matter
knowledge of human experts into a long series of programmed “if given x input,
then provide y output” rules. For example, the AI chess system Deep Blue, which
defeated the world chess champion in 1997, was developed in collaboration
between computer programmers and human chess grandmasters. The
programmers wrote (literally typed by hand) a computer code algorithm that
considered many potential moves and countermoves and reflected rules for
strong chess play given by human experts.
Machine Learning systems are different in that their “knowledge” is not
programmed by humans. Rather, their knowledge is learned from data: a
Machine Learning algorithm runs on a training dataset and produces an AI
model. To a large extent, Machine Learning systems program themselves. Even
so, humans are still critical in guiding this learning process. Humans choose
algorithms, format data, set learning parameters, and troubleshoot problems.
Machine Learning has been around a long time, but it previously was almost
always expensive and complicated with low performance, so there were
comparatively few applications and organizations for which it was a good fit.
Thanks to the ever-increasing availability of massive datasets, massive computing
power (both from using GPU chips as accelerators and from the cloud), open
source code libraries, and software development frameworks, the performance
and practicality of using Machine Learning AI systems has increased dramatically.
There are four different families of Machine Learning algorithms, which differ
based on aspects of the data they train on. It is important to understand the
different families because knowing which family an AI system will use has
implications for effectively enabling and managing the system’s development.

Gregory C. Allen | DoD Joint AI Center 3


Understanding Artificial Intelligence Technology 4

1) Supervised Learning uses example data that has been labeled by human
“supervisors.” Supervised Learning has incredible performance, but getting
sufficient labeled data can be difficult, time-consuming, and expensive.
2) Unsupervised Learning uses data but doesn’t require labels for the data. It has
lower performance than Supervised Learning for many applications, but it
can also be used to tackle problems where Supervised Learning isn’t viable.
3) Semi-Supervised Learning uses both labeled and unlabeled data and has a
mix of the pros and cons of Supervised and Unsupervised learning.
4) Reinforcement Learning has autonomous AI agents that gather their own
data and improve based on their trial and error interaction with the
environment. It shows a lot of promise in basic research, but so far
Reinforcement Learning has been harder to use in the real world. Regardless,
technology firms have many noteworthy, real-world success stories.
Deep Learning (Deep Neural Networks) is a powerful Machine Learning
technique that can be applied to any of the four above families. It provides the
best performance for many applications. However, the technical details are less
important for those not on the engineering staff or directly overseeing the
procurement of these systems. What matters most for program management is
whether or not the system uses Machine Learning, and whether or not the
selected algorithm requires labeled data.
Systems using Machine Learning software can provide very high levels of
performance. However, Machine Learning software has failure modes – both from
accidents and from adversaries – that are distinct from those of traditional
software. Program managers, system developers, test and evaluation personnel,
and system operators all need to be familiar with these failure modes to ensure
safe, secure, and reliable performance of AI systems.
There are multiple steps to developing an operational Machine Learning AI
system. Usually, the biggest challenges relate to getting sufficient high-quality
training data. System performance is directly tied to data quantity, quality, and
representativeness.
Organizations should not pursue using AI for its own sake. Rather, they should have
specific metrics for organizational performance and productivity that they are
seeking to improve. Merely developing a high-performing AI model will not by
itself improve organizational productivity. The model has to be integrated into
operational technology systems, organizational processes, and staff workflows.
Almost always, there will be some changes needed to existing processes to take
full advantage of the AI model’s capability. Adding AI technology without revising
processes will deliver only a tiny fraction of the potential improvements, if any.
Finally, traditional project management wisdom still applies. Many AI projects fail
not because of the technology, but because of a failure to properly set
expectations, integrate with legacy systems, and train operational personnel.

Gregory C. Allen | DoD Joint AI Center 4


Understanding Artificial Intelligence Technology 5

Purpose: By now, nearly all DoD officials understand that the rise of AI is an
important technology trend with significant implications for national security, but
many struggle to give simple and accurate answers to basic questions like:

• What is AI?
• How does AI work?
• Why is now an important time for AI?
• What are the different types of Machine Learning? How do they differ?
• What are Neural Networks and Deep Learning?
• What are the steps of building and operating AI systems?
• What are the limitations and risks of using of AI systems?

Contrary to popular belief, you do not need to understand advanced


mathematics or know computer programming languages to be able to answer
the above questions accurately and to develop a practical understanding of AI
relevant to your organization’s needs. This guide will cover everything that the vast
majority of DoD leaders need to know.

WHAT IS AI?
The DoD AI Strategy states that “AI refers to the ability of machines to perform
tasks that normally require human intelligence.” This definition is so obvious that
many are confused by its simplicity. In fact, however, this definition is very similar
to the ones used by many leading AI textbooks and leading researchers. The first
thing to note about this definition is that AI is an extremely broad field, one that
covers not only the breakthroughs of the past few years, but also the
achievements of the first electronic computers dating back to the 1940s.

The definition of “Artificial Intelligence,” is a bit of a moving target. When


something is new and exciting, people have no qualms about labeling it “Artificial
Intelligence.” Once the capabilities of a particular AI approach are familiar,
though, they are often called merely “software.” This paper will return later to the
subject of how modern AI approaches are different and why now is a critical
moment for AI technology. For now, just understand that even old technology
can still be AI.

HOW DOES AI WORK?


AI is defined by a set of capabilities, rather than a specific technical approach to
achieving those capabilities. There are many different approaches to developing
AI systems, and various approaches work differently with different strengths and
weaknesses. DARPA, a longtime pioneer in AI research, has helpfully grouped
many of these approaches into two broad categories: (1) Handcrafted
Knowledge and (2) Machine Learning. Machine Learning systems are the newer
of the two approaches (though still decades old) and are responsible for the
dramatic improvements in AI capabilities over the past ten years. If you’ve heard

Gregory C. Allen | DoD Joint AI Center 5


Understanding Artificial Intelligence Technology 6

some person or company claim that their system “uses AI,” most likely they mean
that their system is using Machine Learning, which is a far cry from their system
being an autonomous intelligence equal to or greater than human intellect in all
categories. Still, recent progress in Machine Learning is a big deal, with
implications for nearly every industry, including defense and intelligence. The
easiest way to understand Machine Learning systems, however, is by contrasting
them with Handcrafted Knowledge systems, so this paper will begin there.

Handcrafted Knowledge AI
Handcrafted Knowledge Systems are the older of the two AI approaches, nearly
as old as electronic computers. At their core, they are merely software developed
in cooperation between computer programmers and human domain subject
matter experts. Handcrafted Knowledge Systems attempt to represent human
knowledge into programmed sets of rules that computers can use to process
information. In other words, the “intelligence” of the Handcrafted Knowledge
System is merely a very long list of rules in the form of “if given x input, then provide
y output.” When hundreds or thousands or millions of these domain-specific rules
are combined successfully – into “the program” – the result is a machine that can
seem quite smart and can also be very useful.

A well-known example of a Handcrafted Knowledge AI System in widespread use


is tax preparation software. By requiring users to input their tax information
according to pre-specified data formats and then processing that data
according to the formally programmed rules of the tax code (developed in
cooperation between human software engineers and accountants), the output
can be good enough to pass an IRS audit. When first introduced in the 1980s, tax
preparation software was very successfully marketed as Artificial Intelligence.
Now that it has been in widespread use for decades, however, calling it “Artificial
Intelligence” has fallen out of fashion. Nevertheless, it still falls within both the DoD
definition of AI and the formal definition used by most researchers in the field.

Another famous example of a Handcrafted Knowledge System is “Deep Blue,”


the IBM-developed, chess-playing AI that defeated the human world chess
champion in 1997. Deep Blue was developed in cooperation between IBM’s
software engineers and several chess grandmasters, who helped translate their
human chess expertise into tens of thousands of computer code rules for playing
grandmaster-level chess.

Both tax preparation AI systems and Deep Blue are a specific type of Handcrafted
Knowledge AI known as an Expert System. Another type of Handcrafted
Knowledge AI is a Feedback Control System, which uses human-authored rules to
compute system output based on sensor measurement inputs. Feedback Control
Systems have been in widespread use by the Department of Defense for
decades. Aircraft autopilots, missile guidance systems, and electromagnetic

Gregory C. Allen | DoD Joint AI Center 6


Understanding Artificial Intelligence Technology 7

signal processing systems are just a few of the thousands of high-performing,


extremely reliable Feedback Control AI Systems that the Department of Defense
and its partners have developed and operated over the past eight decades. In
this sense, Handcrafted Knowledge Systems are victims of their own success. They
are so common that they are generally no longer referred to as “AI” in common
discourse. Nevertheless, Handcrafted Knowledge Systems remain important and
useful. In some areas, such as tax preparation, they still have far higher
performance than Machine Learning systems. In other areas, such as chess
playing, language translation, and image classification, Handcrafted Knowledge
AI systems have been greatly surpassed in performance by Machine Learning AI
systems. Regardless, Handcrafted Knowledge systems will continue to improve
and see wide use for decades to come.

Machine Learning AI
The key difference between a Handcrafted Knowledge System and a Machine
Learning system is in where it receives its knowledge. Rather than having their
knowledge be provided by humans in the form of hand-programmed rules,
Machine Learning systems generate their own rules. For Machine Learning
systems, humans provide the system training data. By running a human-
generated algorithm on the training dataset, the Machine Learning system
generates the rules such that it can receive input x and provide correct output y.

In other words, the system learns from examples (training data), rather than being
explicitly programmed. This is why data is so vital in the context of AI. Data is the
main raw material out of which high-performing Machine Learning AI systems are
built. For this reason, the quality, quantity, representativeness and diversity of data
will directly impact the operational performance of the ML system. Algorithms and
computing hardware are also important, but nearly all ML systems run on
commodity computing hardware, and nearly all of the best algorithms are freely
available worldwide. Hence, having enough of the right data tends to be the key.

While it is true that – to a large extent – Machine Learning systems program


themselves, humans are still critical in guiding this learning process: humans
choose algorithms and datasets, format data, set learning parameters, and
troubleshoot problems.

At this stage, many readers may ask themselves, “so what? Why is Machine
Learning important?”

The reason is that there are many applications where task automation would be
useful, but where human programming of all of the software rules to implement
automation is either impractical or genuinely impossible. Sometimes human
experts are unable to fully translate their intuition decision-making into fixed rules.
Further, for a surprisingly large subset of these applications, the performance of

Gregory C. Allen | DoD Joint AI Center 7


Understanding Artificial Intelligence Technology 8

Machine Learning systems is very high, much higher than was ever achieved with
Handcrafted Knowledge Systems or indeed by human experts. This does not
mean that Handcrafted Knowledge systems are obsolete. For many applications
they remain the cheapest and/or highest performing approach.

Figure 1: Simplified Diagram of AI Approaches

Pattern recognition, image analysis, language translation, content generation,


and speech transcription are just a few noteworthy examples where the past
performance of Handcrafted Knowledge AI was very low, but the performance
of Machine Learning AI is extremely high, often better than human performance.
Because of the increased performance and enhanced productivity Machine
Learning enables, there are many practical applications throughout the
economy and industry. We are not after using AI for its own sake. We are after
increased performance and enhanced productivity. It’s that simple.

WHY IS NOW AN IMPORTANT TIME FOR AI?


AI has been around for decades. So, why has everyone been talking about it
constantly in recent years? It boils down to this – for Machine Learning AI systems
– there has been a massive increase in the number of real-world applications
where AI is now practical and powerful. There are four main reasons why this is
true now but was not true ten years ago:

Gregory C. Allen | DoD Joint AI Center 8


Understanding Artificial Intelligence Technology 9

1) More Massive Datasets: Machine Learning algorithms tend to require large


quantities of training data in order to produce high performance AI models.
For example, some facial recognition AI systems can now routinely outperform
humans, but to do so requires tens of thousands or millions of labeled images
of faces for training data. When Machine Learning was first developed
decades ago, there were very few applications where sufficiently large
training data was available to build high performance systems. Today, an
enormous number of computers and digital devices and sensors are
connected to the internet, where they are constantly producing and storing
large volumes of data, whether in the form of text, numbers, images, audio, or
other sensor data files.

Of course, more data only helps if the data is relevant to your desired
application. If you’re trying to develop a better aircraft autopilot, then a
bunch of consumer loan application data isn’t going to help, no matter how
much you have. In general, training data needs to match the real-world
operational data very, very closely to train a high-performing AI model.

2) Increased Computing Power: To a far greater extent than Handcrafted


Knowledge Systems, Machine Learning AI systems require a lot of computing
power to process and store all the above-mentioned data. Around ten years
ago, computing hardware started getting powerful enough and cheap
enough that it was possible to run Machine Learning algorithms on massive
datasets using commodity hardware. One especially important turning point
around 2010 was developing effective methods for running Machine Learning
algorithms on Graphics Processing Units (GPUs) rather than on the Central
Processing Units (CPUs) that handle most computing workloads. Originally
designed for video games and computer graphics, GPUs are highly
parallelized, which means they can perform large numbers of similar
calculations at the same time. It turns out that massive parallelism is extremely
useful in speeding up the training of Machine Learning AI models and in
running those models operationally. For many types of Machine Learning,
using GPUs can speed up the training process by 10-20x while reducing
computer hardware costs. Access to the cloud is also very helpful, since
organizations can rapidly access massive computing resources on demand
(for the relatively short amounts of time needed for training) and limit
purchases of computing power to only what they need, when they need it.

3) Improved Machine Learning Algorithms: The first Machine Learning algorithms


are decades old, and some decades-old algorithms remain incredibly useful.
In recent years, however, researchers have discovered many new algorithms
that have greatly sharpened the field’s cutting-edge. These new algorithms
have made Machine Learning models more flexible, more robust, and more
capable of solving different types of problems.

Gregory C. Allen | DoD Joint AI Center 9


Understanding Artificial Intelligence Technology 10

4) Open Source Code Libraries and Frameworks: The cutting-edge of Machine


Learning is not only better than ever, but also more easily available. For a long
time, Machine Learning was a specialized niche within computer science.
Developing Machine Learning systems required a lot of specific expertise and
custom software development that made it out of reach for most
organizations. Now, however, there are many open source code libraries and
developer tools that allow organizations to use and build upon the work of
external communities. As a result, no team or organization has to start from
scratch, and many parts that used to require highly specialized expertise have
been largely automated. The difficulty of developing an AI model has fallen
to the point where – for some applications – even non-experts and beginners
can create useful AI tools. In some cases, open source ML models can be
entirely reused.

Figure 2. Key Factors Driving Recent Improvements in ML Performance

In short, using Machine Learning generally used to be expensive and


complicated, so there were comparatively few applications and organizations
for which it was a good fit. Now, however, using Machine Learning is practical
and powerful for a far more diverse set of applications. Thanks to the ever-
increasing availability of more massive datasets, increased computing power,
improved Machine Learning algorithms, and improved open source code
libraries and software development frameworks, things that used to be nearly
impossible, such as automated facial recognition, are now possible. Programs
that used to have terrible performance, such as automatic translation, now have
significantly better performance. Finally, AI systems that used to be extremely

Gregory C. Allen | DoD Joint AI Center 10


Understanding Artificial Intelligence Technology 11

expensive to develop, such as imagery classification, are often now affordable


and sometimes even cheap.

Despite their huge potential, AI solutions are not a great fit for all types of
problems. If you have an application where you think using AI could be beneficial,
knowing whether or not any particular system that is claiming to use “AI” is using
Machine Learning is important for several reasons. For one thing, Machine
Learning works differently from traditional software, and it has different strengths
and weaknesses too. Moreover, Machine Learning tends to break and fail in
different ways. A basic understanding of these strengths, weaknesses, and failure
modes can help you understand whether or not your particular problems are a
good fit for a Machine Learning AI solution.

WHAT ARE THE DIFFERENT TYPES OF MACHINE LEARNING? HOW DO THEY DIFFER?
Like Artificial Intelligence, Machine Learning is also an umbrella term, and there
are four different broad families of Machine Learning algorithms. There are also
many different subcategories and combinations under these four major families,
but a good understanding of these four broad families will be sufficient for the
vast majority of DoD employees, including senior leaders in non-technical roles.

The four categories – Figure 3. Labeled and Unlabeled Training Data


explained more on
the following page –
differ based on what
types of data their
algorithms can work
with. However, the
important distinction
is not whether the
data is audio,
images, text, or
numbers. Rather, the
important distinction
is whether or not the training data is labeled or unlabeled and how the system
receives its data inputs. Figure 3 provides a simple illustration of labeled and
unlabeled training data for a classifier of images of cats and dogs.

Depending upon whether or not data is labeled, a different family of algorithms


applies. The four major families of algorithms are Supervised Learning,
Unsupervised Learning, Semi-Supervised Learning, and Reinforcement Learning.

Supervised Learning: “Supervised” means that, before the algorithm processes


the training data, some “supervisor” (which may be a human, group of humans,
or a different software system) has accurately labeled each of the data inputs

Gregory C. Allen | DoD Joint AI Center 11


Understanding Artificial Intelligence Technology 12

with its correct associated output. For example, if the goal of the AI system is to
correctly classify the objects in different images as either “cat” or “dog,” the
labeled training data would have image examples paired with the correct
classification label. Supervised Learning systems can also be used for identifying
the correct labels of continuous numerical outputs. For example, “given this wing
shape input, predict the output air drag coefficient.”

Many Supervised Learning systems can achieve extremely high performance, but
they require very large labeled datasets to do so. Using image classification as an
example, a common rule of thumb is that the algorithm needs at least 5,000
labeled examples of each category in order to produce an AI model with decent
performance. Acquiring all of this labeled data can be easy or very difficult,
depending upon the application. In the case of facial recognition algorithms,
most companies use paid humans to manually label images. In the case of online
shopping recommendation engines, the customers are actually providing the
data labels through the normal course of their shopping. The data inputs are the
recommended items displayed to the customers and the customer’s profile
information, while the outputs are the actual purchases made or not made. This
is one of the major reasons why internet companies were at the forefront of the
adoption of Machine Learning AI: their users were constantly producing valuable
datasets – both labeled and unlabeled – and the online environment allowed for
rapid experimentation with Machine Learning-enabled analysis and automation.

Note that pre-labeled data is only required for the training data that the algorithm
uses to train the AI model. The AI model in operational use with new data will be
generating its own labels, the accuracy of which will depend on the AI’s training.
If the training data set was sufficiently large, high quality, and representative of
the diversity present in the operational environment, then the performance of the
AI model in generating these labels can be at or above human performance.

Unsupervised Learning: Unsupervised algorithms are those that can extract


features from the data without the need for a ground-truth label for the results.
Using the aforementioned example of an image classifier, the AI model produced
by an unsupervised algorithm would not return that a specific input image was of
a “cat” or a “dog.” Rather, the model would sort the training dataset into various
groups based on their similarity. One sorted group might be the desired groups of
cats and dogs, but images might instead be sorted based on undesired
categories such as whether or not they have a blue sky in the background, or a
wooden floor. Unsupervised Learning systems are therefore often less predictable,
but because unlabeled data is almost always more available than labeled data,
they remain critical. Additionally, Unsupervised algorithms are very useful when
developers seek to explore and understand their own datasets and what
properties might be useful in either developing automation or changing
operational practices and policies.

Gregory C. Allen | DoD Joint AI Center 12


Understanding Artificial Intelligence Technology 13

Figure 4 depicts the differences between Supervised and Unsupervised algorithms


using an image analysis example.

Figure 4. Illustrated Example of Supervised and Unsupervised Algorithms

It is not true that Unsupervised Learning is “worse” than Supervised Learning


(though performance can be lower for some use cases). Rather, Unsupervised
Learning is useful for solving different types of problems. A common Unsupervised
use case is fraud detection in financial data. In the case of fraud detection,
Supervised Learning could be a good fit for identifying potential fraud that
matches behaviors known to be unlawful or associated with fraud. Unsupervised
Learning can find new, unidentified patterns of behavior that might indicate new
types of fraud techniques.

Gregory C. Allen | DoD Joint AI Center 13


Understanding Artificial Intelligence Technology 14

Semi-Supervised Learning: There is also an increasingly popular class of “Semi-


Supervised” algorithms that combine techniques from Supervised and
Unsupervised algorithms for applications with a small set of labeled data and a
large set of unlabeled data. In practice, using them leads to exactly what you
would expect, a mix of some of both of the strengths and weaknesses of
Supervised and Unsupervised approaches.

Reinforcement Learning: In Reinforcement Learning, the training data is collected


by an autonomous, self-directed AI agent in the course of perceiving its
environment (which might be the real world or a simulated environment) and
performing goal-directed actions (trying to maximize receipt of “rewards”). Four
aspects of Reinforcement Learning are notably distinct from Supervised and
Unsupervised Learning:

1) Data is gathered by the AI agent itself in the course of its interacting with
the environment and perceiving stated changes. For example, an AI agent
playing a digital game of chess makes moves and perceives changes in
the board based on its moves.
2) The rewards are input data received by the agent when certain criteria are
satisfied. For example, a Reinforcement Learning AI agent in chess will
make many moves before each win or loss. These criteria are typically
unknown to the agent at the outset of training.
3) Rewards often contain only partial information. A reward like a win in chess
conveys that some inputs must have been good, but it doesn’t clearly
signal which inputs were good and which were not.
4) The system is learning an action policy for taking actions to maximize its
receipt of cumulative rewards.

Figure 5: Simplified Reinforcement Learning Diagram

Gregory C. Allen | DoD Joint AI Center 14


Understanding Artificial Intelligence Technology 15

Reinforcement Learning has proven to be very useful in developing high


performance AI systems that play games, such as chess, Go, poker, and
StarCraft II, including systems that can defeat human world-champion-level
players. In developing learning systems for these games, there are no labeled
datasets that outline every possible move that can be made and provide a true
assessment of whether it was a “good move” or a “bad move.” Instead, the
partial labels only reveal that the final outcome of all the moves made in the
game was a “win” or a “loss.” Reinforcement algorithms explore the space of
possible actions in an effort to learn the optimal policy (set of rules for determining
the best action) that will maximize long term rewards.

Reinforcement Learning works very well for games and simulations because the
system automatically generates its own training data, which only costs the price
of running computational hardware for the algorithm and simulation. For
example, AlphaGo, a Reinforcement Learning system focused on the board
game Go, played more than 4.9 million games in three days (one full game every
nineteen seconds) against itself in order to learn how to play the game at a world-
champion level. In real life, a Go game takes ~1 hour.

Reinforcement Learning is currently more challenging to utilize in applications that


operate in the real world for three reasons. First, the real world is not as heavily
bounded as video games in terms of inputs, outputs, and interactions. Second,
time cannot be sped up in the real world. Third, there are consequences to failure
in the real world. For example, nothing bad happens when an autonomous
driving system crashes in simulation. In the real world, the consequences can be
severe, even lethal. As a practical matter, Reinforcement Learning systems have
shown significant promise in research settings, but they are much less common
than Supervised and Unsupervised Learning applications in real world operations.
Nevertheless, many tech companies do use Reinforcement Learning in
operational applications, and many researchers expect that real-world usage will
grow significantly over the next decade.

Different Types of Machine Learning Summarized


Knowing the different families of Machine Learning algorithms and their strengths
and weaknesses is fundamental to making wise decisions about developing and
using AI systems. Supervised Learning systems can deliver incredible performance,
but acquiring labeled data may be a challenge. Unsupervised Learning systems
don’t require labeled data, but their performance for some applications will
generally be far more limited than Supervised systems. Reinforcement Learning
systems can generate their own data but can generally only be used for
applications that offer access to simulators that closely resemble the operational
environment. Thus far, this includes fewer applications.

Gregory C. Allen | DoD Joint AI Center 15


Understanding Artificial Intelligence Technology 16

When pursuing a specific AI application, developers and program managers


should always start by asking the following questions:

• How much training data do we have that is relevant to our desired


application? How can we get more?
• How accurate and consistent is the data? How can we increase this?
• Is this data labeled? If not, how difficult would it be to label this data?
• Can only highly trained human experts label the data, or can ordinary
people label it?
• Is our training data truly representative of and appropriate for the
operational environment? How will this change over time?
• After operational deployment, what types of data changes could cause
this system to have degraded performance?
• Do we have access to a high-fidelity simulator that closely resembles our
operational environment?

Particular software or hardware products are usually integrated systems with


many different elements and functions. One part of a system might use
Supervised Learning, while another might use Unsupervised, and so on. In each
case, however, those questions will guide the development and management of
each subsystem or element of the integrated system.

WHAT ARE NEURAL NETWORKS AND MACHINE LEARNING?


In the mass media, Neural Networks and Deep Learning are two of the most
common phrases associated with excitement about modern AI. There’s a good
reason for that: the dramatic improvements in the performance of AI systems,
especially those of recent research breakthroughs, have for the most part been
enabled by algorithms that make use of Neural Networks.

Neural Networks are a specific category of algorithms that are very loosely
inspired by biological neurons in the brain. Deep Neural Networks (a.k.a. Deep
Learning) merely refers to those Neural Networks that have many layers of
connected neurons in sequence (“deep” referring to the number of layers).
Though Neural Networks are most strongly associated with Supervised Learning,
Deep Learning can, with the right architecture, also be applied to Unsupervised,
Semi-Supervised, and Reinforcement Learning.

Neural Networks have been around since the late 1950s, but training Deep Neural
Networks only became practical around the 2006 timeframe. Since then, the
previously mentioned trends – more data, more computing power, improved
algorithms, and improved open source code libraries – have had an especially
large impact on improving Deep Learning performance. Since 2012, many of the
winning systems in AI competitions around important performance benchmarks

Gregory C. Allen | DoD Joint AI Center 16


Understanding Artificial Intelligence Technology 17

are routinely won by systems that make use of Neural Networks and Deep
Learning.

While it is very important for engineers and developers of DoD AI systems and DoD
technical leaders to have a good understanding of Neural Networks and how
they work, a granular understanding of Neural Networks is overkill and a
distraction for most DoD senior leaders. Everything stated earlier in this paper
about the general categories of Machine Learning and different types of
Machine Learning applies to Neural Networks as well. And knowing whether or
not your Machine Learning system is using Neural Networks or another algorithm
like decision trees won’t have many important implications for how you run your
program. In general, AI program managers should care about the performance
of the AI system and the types of data required in order to guarantee that
performance. For many applications, achieving the required performance will
require using Neural Networks. For others, it won’t. Focusing on using the most
advanced algorithm is important for many parts of the basic research community.
For those in positions involving policy, applied R&D, and operations, factors like
feasibility, performance, and reliability are more important.

Figure 6. Deep Learning’s Place in AI – Using the DARPA “AI Waves” Framework

There is one important exception, however. Neural Networks differ from other
types of Machine Learning algorithms in that they tend to have low explainability.
The system can generate a prediction or other output, and testing can provide
evidence suggesting that these predictions have high accuracy, but it is very
difficult to understand or explain the specific causal mechanisms by which the
Neural Network arrived at its prediction, even for top AI experts. This “explainability
problem” is often described as a problem for all of AI, but it is primarily a problem
for Neural Networks and Deep Learning. Many other types of Machine Learning
algorithms – for example decision trees – have very high explainability.

Gregory C. Allen | DoD Joint AI Center 17


Understanding Artificial Intelligence Technology 18

WHAT ARE THE STEPS OF BUILDING AND OPERATING MACHINE LEARNING SYSTEMS?
By now you should have a good understanding of what the different families of
Machine Learning algorithms are and how they work. Figure 7 shows the major
steps of actually generating Machine Learning models for operational use.

The ultimate goal of an AI development effort is an AI Model that delivers good


performance for a given application at acceptable cost. An AI model built with
Machine Learning is generated by providing prepared training data to a suitable
AI algorithm. The vast majority of tasks during the AI model training process are
performed automatically by the algorithm, but in almost all cases this process will
need to be overseen and calibrated by human Machine Learning experts.

Figure 7: Simplified Depiction of the Machine Learning Development Process

The AI model will also likely have to be integrated with existing systems and pass
through a suitable testing and evaluation process. Having a high performing AI
model by itself, however, is not enough to deliver a positive impact on
organizational productivity. The most significant organizational productivity
enhancements require not just enhanced technical performance, but also
operational processes and staff workflow changes that effectively take
advantage of the enhanced performance.

WHAT ARE THE LIMITATIONS OF MACHINE LEARNING SYSTEMS?


AI systems are subject to failures resulting both from accidents (safety failures) and
from adversarial malicious activity (security failures). There are many different
types of Machine Learning failure modes, but perhaps the most common is when
the training data is not sufficiently representative and instructive for the diverse,
real-world examples the Machine Learning system will encounter. For example, a
satellite imagery classifier that is trained to recognize vehicles exclusively using
training data images in a desert environment should be assumed to have
degraded performance if the operational data images are of the same vehicles
in a grassland, urban, or arctic tundra environment. For the same reason, the
performance of ML models in real world applications generally degrades over
time if not regularly updated with new training data that reflects the changing
state of the world. The software engineering maxim “software is never done” is
doubly true for Machine Learning software.

Gregory C. Allen | DoD Joint AI Center 18


Understanding Artificial Intelligence Technology 19

More broadly, there are not yet widely agreed upon safety and reliability
standards for the development, testing, and operation of Machine Learning
systems. Some methods that have proved critical in ensuring safety and reliability
of traditional software – such as formal verification – are not currently available
for use on Machine Learning systems. Moreover, some Machine Learning failure
modes are not fully understood even at the basic research level. Despite the
current challenges, Machine Learning AI systems are already (for some cases)
safer and better performing than what they replace. With additional future R&D
and improved program management standards, Machine Learning will also be
reliably used in a much more diverse set of applications, including safety-critical
ones. When this positive scenario is realized, it will not be because AI systems are
inherently safe and secure – no technology is – but because the responsible
stakeholders took the necessary steps to make AI safe and secure. For the DoD,
there are very promising developments on this issue. In February 2020, the DoD
officially adopted five Ethical Principles for Artificial Intelligence. The DoD also
established an Executive Steering Group with representation from each of the
Armed Services and major DoD components to make recommendations for
improvements in all aspects of DoD AI Policy and operational usage.

CONCLUSION
Talented human capital and access to AI experts are critical factors for success
in DoD’s AI strategy. However, the basics of AI technology can be understood by
anyone who devotes the time to learn. The concepts in this document provide a
technical overview that will be adequate for the vast majority of senior leaders to
understand what would be required to adopt and utilize AI for their organization.

Those who want to go further and learn more are encouraged to do so. This
document can serve as a useful jumping off point to more advanced and
domain-specific subjects. Some recommendations for further reading are
provided on the next page. Before moving on, however, readers would be wise
to double down and ensure they have a rock-solid understanding of the basics.

Gregory C. Allen | DoD Joint AI Center 19


Understanding Artificial Intelligence Technology 20

REFERENCES AND RECOMMENDED FURTHER READING


For non-technical individuals seeking to better understand Machine Learning
• Launchbury, J. (2017, March 19). A DARPA Perspective on Artificial Intelligence.
Retrieved from https://machinelearning.technicacuriosa.com/2017/03/19/a-darpa-
perspective-on-artificial-intelligence/
• Karpathy, A. (2017, November 11). Software 2.0. Retrieved from
https://medium.com/@karpathy/software-2-0-a64152b37c35
• Domingos, P. (2016). The master algorithm: how the quest for the ultimate learning
machine will remake our world. New York: Basic books

For engineers seeking to better understand Machine Learning


• Lecun, Y., Bengio, Y., & Hinton, G. (2015). Deep Learning. Nature, 521(7553), 436–444.
doi: 10.1038/nature14539
• Domingos, P. (2012). A few useful things to know about Machine Learning.
Communications of the ACM, 55(10), 78–87. doi: 10.1145/2347736.2347755
• Brunton, S., Noack, B., & Koumoutsakos, P. (2020, January 4). Machine Learning for Fluid
Mechanics. Retrieved from https://arxiv.org/abs/1905.11075
• Russell, S. J., & Norvig, P. (2010). Artificial Intelligence: a modern approach (3rd ed.).
Upper Saddle River: Prentice-Hall. [note: 4th edition expected in mid-2020]
• Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. Cambridge, MA: The
MIT Press.

Helpful context for reading popular media about AI


• Karpathy, A. (2017, May 31). AlphaGo, in context. Retrieved from
https://medium.com/@karpathy/alphago-in-context-c47718cb95a5

Organizational best practices for using AI to improve performance


• Berinato, S. (2017, July 19). Inside Facebook's AI Workshop. Retrieved from
https://hbr.org/2017/07/inside-facebooks-ai-workshop
• Horneman, A., Mellinger, A., & Ozkaya, I. (2019). AI Engineering: 11 Foundational
Practices. Carnegie Mellon University, Software Engineering Institute.
https://resources.sei.cmu.edu/asset_files/WhitePaper/2019_019_001_634648.pdf
• Basilico, J. (2017, December 13). Making Netflix Machine Learning Algorithms Reliable.
Retrieved from https://www.slideshare.net/justinbasilico/making-netflix-machine-
learning-algorithms-reliable
• Kim, G., Debois, P., Willis, J., Humble, J., & Allspaw, J. (2016). The DevOps handbook
Portland, OR: IT Revolution Press, LLC.

Further Reading on AI Safety and Security


• Department of Defense (2020). DOD Adopts Ethical Principles for Artificial Intelligence.
https://www.defense.gov/Newsroom/Releases/Release/Article/2091996/dod-adopts-
ethical-principles-for-artificial-intelligence/
• Ortega, P. A., & Maini, V. (2018, September 27). DeepMind Safety Research: Building
safe Artificial Intelligence https://medium.com/@deepmindsafetyresearch/building-
safe-artificial-intelligence-52f5f75058f1
• Amodei, D., Olah, C., Steinhardt, J., Christiano, P., & Dan. (2016, July 25). Concrete
Problems in AI Safety. Retrieved from https://arxiv.org/abs/1606.06565
• Nguyen, A., Yosinski, J., & Clune, J. (2015). Deep Neural Networks are easily fooled: High
confidence predictions for unrecognizable images. 2015 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR). doi: 10.1109/cvpr.2015.7298640

Gregory C. Allen | DoD Joint AI Center 20

You might also like