Artificial Intelligence/

Machine Learning Explained

Author: Steve Blank

Gordian Knot Center for

National Security Innovation

https://gordianknot.stanford.edu
Artificial Intelligence/Machine Learning– Explained

AI is a once-in-a lifetime commercial and defense game changer

Hundreds of billions in public and private capital is being invested in AI and Machine Learning
companies. The number of patents filed in 2021 is more than 30 times higher than in 2015 as
companies and countries across the world have realized that AI and Machine Learning will be a
major disruptor and potentially change the balance of military power.

Until recently, the hype exceeded reality. Today, however, advances in AI in several important
areas (here, here, here, here and here) equal and even surpass human capabilities.

If you haven’t paid attention, now’s the time.

AI and the DoD

The Department of Defense has thought that AI is such a foundational set of technologies that
they started a dedicated organization- the JAIC - to enable and implement artificial intelligence
across the Department. They provide the infrastructure, tools, and technical expertise for DoD
users to successfully build and deploy their AI-accelerated projects.

Some specific defense related AI applications are listed later in this document.

We’re in the Middle of a Revolution

Imagine it’s 1950, and you’re a visitor who traveled back in time from today. Your job is to
explain the impact computers will have on business, defense and society to people who are
using manual calculators and slide rules. You succeed in convincing one company and a
government to adopt computers and learn to code much faster than their competitors
/adversaries. And they figure out how they could digitally enable their business – supply chain,
customer interactions, etc. Think about the competitive edge they’d have by today in business
or as a nation. They’d steamroll everyone.

That’s where we are today with Artificial Intelligence and Machine Learning. These technologies
will transform businesses and government agencies. Today, 100s of billions of dollars in private
capital have been invested in 1,000s of AI startups. The U.S. Department of Defense has created
a dedicated organization to ensure its deployment.

But What Is It?

Compared to the classic computing we’ve had for the last 75 years, AI has led to new types of
applications, e.g. facial recognition; new types of algorithms, e.g. machine learning; new types
of computer architectures, e.g. neural nets; new hardware, e.g. GPUs; new types of software
developers, e.g. data scientists; all under the overarching theme of artificial intelligence. The
sum of these feels like buzzword bingo. But they herald a sea change in what computers are
capable of doing, how they do it, and what hardware and software is needed to do it.
This brief will attempt to describe all of it.

New Words to Define Old Things

One of the reasons the world of AI/ML is confusing is that it’s created its own language and
vocabulary. It uses new words to define programming steps, job descriptions, development
tools, etc. But once you understand how the new world maps onto the classic computing world,
it starts to make sense. So first a short list of some key definitions.

AI/ML - a shorthand for Artificial Intelligence/Machine Learning

Artificial Intelligence (AI) - a catchall term used to describe “Intelligent machines” which can
solve problems, make/suggest decisions and perform tasks that have traditionally required
humans to do. AI is not a single thing, but a constellation of different technologies.

Machine Learning (ML) - a subfield of

artificial intelligence. Humans
combine data with algorithms (see
here for a list) to train a model using
that data. This trained model can
then make predications on new data
(is this picture a cat, a dog or a
person?) or decision-making
processes (like understanding text
and images) without being explicitly
programmed to do so.

Machine learning algorithms - computer programs that adjust themselves to perform

better as they are exposed to more data. The “learning” part of machine learning means
these programs change how they process data over time. In other words, a machine-
learning algorithm can adjust its own settings, given feedback on its previous
performance in making predictions about a collection of data (images, text, etc.).

Deep Learning/Neural Nets – a subfield of machine learning. Neural networks make up the
backbone of deep learning. (The “deep” in deep learning
refers to the depth of layers in a neural network.) Neural
nets are effective at a variety of tasks (e.g., image
classification, speech recognition). A deep learning neural
net algorithm is given massive volumes of data, and a task to
perform - such as classification. The resulting model is
capable of solving complex tasks such as recognizing objects
within an image and translating speech in real time. In
reality, the neural net is a logical concept that gets mapped
onto a physical set of specialized processors. See here.)
Data Science – a new field of computer science. Broadly it encompasses data systems and
processes aimed at maintaining data sets and deriving meaning out of them. In the context of
AI, it’s the practice of people who are doing machine learning.

Data Scientists - responsible for extracting insights that help businesses make decisions.
They explore and analyze data using machine learning platforms to create models about
customers, processes, risks, or whatever they’re trying to predict.

What’s Different? Why is Machine Learning Possible Now?

To understand why AI/Machine Learning can do these things, let’s compare them to computers
before AI came on the scene. (Warning – simplified examples below.)

Classic Computers
For the last 75 years computers (we’ll call these classic computers) have both shrunk to pocket
size (iPhones) and grown to the size of warehouses (cloud data centers), yet they all continued
to operate essentially the same way.

Classic Computers - Programming

Classic computers are designed to do anything a human explicitly tells them to do. People
(programmers) write software code (programming) to develop applications, thinking a priori
about all the rules, logic and knowledge that need to be built in to an application so that it can
deliver a specific result. These rules are explicitly coded into a program using a software
language (Python, JavaScript, C#, Rust, …).

Classic Computers - Compiling

The code is then compiled using software to translate the programmer’s source code into a
version that can be run on a target computer/browser/phone. For most of today’s programs,
the computer used to develop and compile the code does not have to be that much faster than
the one that will run it.
Classic Computers - Running/Executing Programs
Once a program is coded and compiled, it can be deployed and run (executed) on a desktop
computer, phone, in a browser window, a data center cluster, in special hardware, etc.
Programs/applications can be games, social media, office applications, missile guidance
systems, bitcoin mining, or even operating systems e.g. Linux, Windows, IOS. These programs
run on the same type of classic computer architectures they were programmed in.

Classic Computers – Software Updates, New Features

For programs written for classic computers, software developers receive bug reports, monitor
for security breaches, and send out regular software updates that fix bugs, increase
performance and at times add new features.

Classic Computers- Hardware

The CPUs (Central Processing Units) that write and run these Classic Computer applications all
have the same basic design (architecture). The CPUs are designed to handle a wide range
of tasks quickly in a serial fashion. These CPUs range from Intel X86 chips, and the ARM cores
on Apple M1 SoC, to the z15 in IBM mainframes.

Machine Learning
In contrast to programming on classic computing with fixed rules, machine learning is just like it
sounds – we can train/teach a computer to “learn by example” by feeding it lots and lots of
examples. (For images a rule of thumb is that a machine learning algorithm needs at least 5,000
labeled examples of each category in order to produce an AI model with decent performance.)
Once it is trained, the computer runs on its own and can make predictions and/or complex
decisions.

Just as traditional programming has three steps - first coding a program, next compiling it and
then running it - machine learning also has three steps: training (teaching), pruning and
inference (predicting by itself.)
Machine Learning - Training
Unlike programing classic computers with explicit rules, training is the process of “teaching” a
computer to perform a task e.g. recognize faces, signals, understand text, etc. (Now you know
why you're asked to click on images of traffic lights, cross walks, stop signs, and buses or type
the text of scanned image in ReCaptcha.) Humans provide massive volumes of “training data”
(the more data, the better the model’s performance) and select the appropriate algorithm to
find the best optimized outcome.
(See the detailed “machine learning pipeline” later in this section for the gory details.)

By running an algorithm selected by a data scientist on a set of training data, the Machine
Learning system generates the rules embedded in a trained model. The system learns from
examples (training data), rather than being explicitly programmed. (See the “Types of Machine
Learning” section for more detail.) This self-correction is pretty cool. An input to a neural net
results in a guess about what that input is. The neural net then takes its guess and compares it
to a ground-truth about the data, effectively asking an expert “Did I get this right?” The
difference between the network’s guess and the ground truth is its error. The network
measures that error, and walks the error back over its model, adjusting weights to the extent
that they contributed to the error.)

Just to make the point again: The algorithms combined with the training data - not external
human computer programmers - create the rules that the AI uses. The resulting model is
capable of solving complex tasks such as recognizing objects it’s never seen before, translating
text or speech, or controlling a drone swarm.
(Instead of building a model from scratch you can now buy, for common machine learning
tasks, pretrained models from others and here, much like chip designers buying IP Cores.)

Machine Learning Training - Hardware

Training a machine learning model is a very computationally intensive task. AI hardware must
be able to perform thousands of multiplications and additions in a mathematical process called
matrix multiplication. It requires specialized chips to run fast. (See the AI hardware section for
details.)

Machine Learning - Simplification via pruning, quantization, distillation

Just like classic computer code needs to be compiled and optimized before it is deployed on its
target hardware, the machine learning models are simplified and modified (pruned) to use less
computing power, energy, and memory before they’re deployed to run on their hardware.

Machine Learning – Inference Phase

Once the system has been trained it can be copied to other devices and run. And the computing
hardware can now make inferences (predictions) on new data that the model has never
seen before.

Inference can even occur locally on edge devices where physical devices meet the digital world
(routers, sensors, IOT devices), close to the source of where the data is generated. This reduces
network bandwidth issues and eliminates latency issues.

Machine Learning Inference - Hardware

Inference (running the model) requires substantially less compute power than training. But
inference also benefits from specialized AI chips.

Machine Learning – Performance Monitoring and Retraining

Just like classic computers where software developers do regular software updates to fix bugs
and increase performance and add features, machine learning models also need to be updated
regularly by adding new data to the old training pipelines and running them again. Why?
Over time machine learning models get stale. Their real-world performance generally degrades
over time if they are not updated regularly with new training data that matches the changing
state of the world. The models need to be monitored and retrained regularly for data and/or
concept drift, harmful predictions, performance drops, etc. To stay up to date, the models need
to re-learn the patterns by looking at the most recent data that better reflects reality.

One Last Thing – “Verifiability/Explainability”

Understanding how an AI works is essential to fostering trust and confidence in AI production
models.

Neural Networks and Deep Learning differ from other types of Machine Learning algorithms in
that they have low explainability. They can generate a prediction, but it is very difficult to
understand or explain how it arrived at its prediction. This “explainability problem” is often
described as a problem for all of AI, but it’s primarily a problem for Neural Networks and Deep
Learning. Other types of Machine Learning algorithms – for example decision trees – have very
high explainability. The results of the five-year DARPA Explainable AI Program (XAI) are worth
reading here.

So What Can Machine Learning Do?1

It’s taken decades but as of today, on its simplest implementations, machine learning
applications can do some tasks better and/or faster than humans. Machine Learning is most
advanced and widely applied today in processing text (through Natural Language Processing)

1
https://databricks.com/discover/pages/the-democratization-of-artificial-intelligence-and-deep-learning
followed by understanding images and videos (through Computer Vision) and analytics and
anomaly detection. For example:

Recognize and Understand Text/Natural Language Processing

AI is better than humans on basic reading comprehension benchmarks like SuperGLUE and
SQuAD and their performance on complex linguistic tasks is almost there. Applications: GPT-3,
M6, OPT-175B, Google Translate, Gmail Autocomplete, Chatbots, Text summarization.

Write Human-like Answers to Questions and Assist in Writing Computer Code

An AI can write original text that is indistinguishable from that created by humans. Examples
GPT-3, Wu Dao 2.0 or generate computer code. Example GitHub Copilot, Wordtune

Recognize and Understand Images and video streams

An AI can see and understand what it sees. It can identify and detect an object or a
feature in an image or video. It can even identify faces. It can scan news broadcasts
or read and assess text that appears in videos. It has uses in threat detection -
airport security, banks, and sporting events. In medicine to interpret MRI’s or to
design drugs. And in retail to scan and analyze in-store imagery to intuitively determine
inventory movement. Examples of ImageNet benchmarks here and here

Detect Changes in Patterns/Recognize Anomalies

An AI can recognize patterns which don’t match the behaviors expected for a
particular system, out of millions of different inputs or transactions. These
applications can discover evidence of an attack on financial networks, fraud
detection in insurance filings or credit card purchases; identify fake reviews; even tag
sensor data in industrial facilities that mean there’s a safety issue. Examples here, here and
here.

Power Recommendation Engines

An AI can provide recommendations based on user behaviors used in ecommerce
to provide accurate suggestions of products to users for future purchases based on
their shopping history. Examples: Alexa and Siri

Recognize and Understand Your Voice

An AI can understand spoken language. Then it can comprehend what is being said and in what
context. This can enable chatbots to have a conversation with
people. It can record and transcribe meetings. (Some versions can
even read lips to increase accuracy.) Examples: Siri/Alexa/Google
Assistant. Example here

Create Artificial Images

AI can create artificial images (DeepFakes) that are
indistinguishable from real ones using Generative Adversarial
Networks. Useful in entertainment, virtual worlds, gaming, fashion design, etc. Synthetic faces
are now indistinguishable and more trustworthy than photos of real people. Paper here.

Create Artist Quality Illustrations from A Written Description

AI can generate images from text descriptions, creating anthropomorphized versions of animals
and objects, combining unrelated concepts in plausible ways. An example is Dall-E

Generative Design of Physical Products

Engineers can input design goals into AI-driven generative design software, along with
parameters such as performance or spatial requirements, materials, manufacturing methods,
and cost constraints. The software explores all the possible permutations of a solution, quickly
generating design alternatives Example here.

Sentiment Analysis
An AI leverages deep natural language processing, text analysis, and computational linguistics
to gain insight into customer opinion, understanding of consumer sentiment, and
measuring the impact of marketing strategies. Examples: Brand24, MonkeyLearn

What Does this Mean for Businesses?

Skip this section if you’re interested in national security applications

Hang on to your seat. We’re just at the beginning of the revolution. The next phase of AI,
powered by ever increasing powerful AI hardware and cloud clusters, will combine some of
these basic algorithms into applications that do things no human can. It will transform business
and defense in ways that will create new applications and opportunities.

Human-Machine Teaming
Applications with embedded intelligence have already begun to appear thanks to massive
language models. For example - Copilot as a pair-programmer in Microsoft Visual Studio
VSCode. It’s not hard to imagine DALL-E 2 as an illustration assistant in a photo editing
application, or GPT-3 as a writing assistant in Google Docs.

AI in Medicine
AI applications are already appearing in radiology, dermatology, and oncology. Examples: IDx-
DR, OsteoDetect, Embrace2. AI Medical image identification can automatically detect lesions,
and tumors with diagnostics equal to or greater than humans. For Pharma, AI will power drug
discovery design for finding new drug candidates. The FDA has a plan for approving AI software
here has a list of AI-enabled medical devices here.
Autonomous Vehicles
Harder than it first seemed, but car companies like Tesla will eventually get better than human
autonomy for highway driving and eventually city streets.

Decision support
Advanced virtual assistants can listen to and observe behaviors, build and maintain data
models, and predict and recommend actions to assist people with and automate tasks that
were previously only possible for humans to accomplish.

Supply chain management

AI applications are already appearing in predictive maintenance, risk management,
procurement, order fulfillment, supply chain planning and promotion management.

Marketing
AI applications are already appearing in real-time personalization, content and media
optimization and campaign orchestration to augment, streamline and automate marketing
processes and tasks constrained by human costs and capability, and to uncover new customer
insights and accelerate deployment at scale.

Making business smarter: Customer Support

AI applications are already appearing in virtual customer assistants with speech recognition,
sentiment analysis, automated/augmented quality assurance and other technologies providing
customers with 24/7 self- and assisted-service options across channels.

AI in National Security2
Much like the dual-use/dual-nature of classical computers AI developed for commercial
applications can also be used for national security.

AI/ML and Ubiquitous Technical Surveillance

AI/ML have made most cities untenable for traditional
tradecraft. Machine learning can integrate travel data
(customs, airline, train, car rental, hotel, license plate
readers…,) integrate feeds from CCTV cameras for facial
recognition and gait recognition, breadcrumbs from wireless
devices and then combine it with DNA sampling. The result
is automated persistent surveillance.

China’s employment of AI as a tool of repression and

surveillance of the Uyghurs is a dystopian of how a totalitarian regimes will use AI-enable
ubiquitous surveillance to repress and monitor its own populace.

2
https://www.nscai.gov/2021-final-report/
AI/ML on the Battlefield
AI will enable new levels of performance and autonomy for weapon systems. Autonomously
collaborating assets (e.g., drone swarms, ground vehicles) that can coordinate attacks, ISR
missions, & more.

Fusing and making sense of sensor data (detecting threats in optical /SAR imagery, classifying
aircraft based on radar returns, searching for anomalies in radio frequency signatures, etc.)
Machine learning is better and faster than humans in finding targets hidden in a high-clutter
background. Automated target detection and fires from satellite/UAV.

For example, an Unmanned Aerial Vehicle (UAV) or Unmanned Ground Vehicles with on board
AI edge computers could use deep learning to detect and locate concealed chemical, biological
and explosives threats by fusing imaging sensors and chemical/biological sensors.
Other examples include:

Use AI/ML countermeasures against adversarial, low probability of intercept/low probability of

detection (LPI/LPD) radar techniques in radar and communication systems.

Given sequences of observations of unknown radar waveforms from arbitrary emitters without
a priori knowledge, use machine learning to develop behavioral models to enable inference of
radar intent and threat level, and to enable prediction of future behaviors.

For objects in space, use machine learning to predict and characterize a spacecrafts possible
actions, its subsequent trajectory, and what threats it can pose from along that trajectory.
Predict the outcomes of finite burn, continuous thrust, and impulsive maneuvers.

AI empowers other applications such as:

• Flight Operations Planning Decision Aid Tool for Strike Operations Aboard Aircraft
Carriers
• Automated Battle management – air and missile defense, army/navy tactical…

AI/ML in Collection
The front end of intelligence collection platforms has created a firehose of data that have
overwhelmed human analysts. “Smart” sensors coupled with inference engines that can pre-
process raw intelligence and prioritize what data to transmit and store –helpful in degraded or
low-bandwidth environments.

Human-Machine Teaming in Signals Intelligence

Applications with embedded intelligence have already begun to appear in commercial
applications thanks to massive language models. For example - Copilot as a pair-programmer in
Microsoft Visual Studio VSCode. It’s not hard to imagine an AI
that can detect and isolate anomalies and other patterns of
interest in all sorts of signal data faster and more reliably
than human operators.

AI-enabled natural language processing, computer vision,

and audiovisual analysis can vastly reduce manual data
processing. Advances in speech-to-text transcription and
language analytics now enable reading comprehension,
question answering, and automated summarization of large quantities of text. This not only
prioritizes the work of human analysts, it’s a major force multiplier

AI can also be used to automate data conversion such as translations and decryptions,
accelerating the ability to derive actionable insights.

Human-Machine Teaming in Tasking and Dissemination

AI-enabled systems will automate and optimize tasking and collection for platforms, sensors,
and assets in near-real time in response to dynamic intelligence requirements or changes in the
environment.

AI will be able to automatically generate machine-readable versions of intelligence products

and disseminate them at machine speed so that computer systems across the IC and the
military can ingest and use them in real time without manual intervention.

Human-Machine Teaming in Exploitation and Analytics

AI-enabled tools can augment filtering, flagging, and triage across multiple data sets. They can
identify connections and correlations more efficiently and at a greater scale than human
analysts, and can flag those findings and the most important content for human analysis.
AI can fuse data from multiple sources, types of intelligence, and classification levels to produce
accurate predictive analysis in a way that is not currently possible. This can improve indications
and warnings for military operations and active cyber defense.

AI/ML Information warfare

Nation states have used AI systems to enhance disinformation campaigns and cyberattacks.
This included using “DeepFakes” (fake videos generated by a neural network that are nearly
indistinguishable from reality). They are harvesting data on Americans to build profiles of our
beliefs, behavior, and biological makeup for tailored attempts to manipulate or coerce
individuals.

But because a large percentage of it is open-source AI is not limited to nation states, AI-
powered cyber-attacks, deepfakes and AI software paired with commercially available drones
can create “poor-man’s smart weapons” for use by rogue states, terrorists and criminals.
 closer to home. Because of AI, adversaries will be able to act with
micro-precision, but at macro-scale and with greater speed. They will

and to target individuals in new ways. AI will also help create precisely
Graphic 1.1: How AI is Transforming the Threat Landscape
AI/ML Cyberwarfarebiological agents. And adversaries will manipulate the AI
engineered
AI-enabled malware can learn and adapt to a system’s defensive measures, or, conversely, AI-
systems we will rely upon.
enabled cyber-defensive tools can proactively locate and address network anomalies and
system vulnerabilities.
AI – Friend or Foe?
Current Threats New Threats Threats TO AI Stacks Future Threats
Advanced BY AI Systems FROM AI Systems Themselves VIA AI Systems
Transforming the
hreat L andscape AI transforms existing AI creates new AI itself is also a new Examples of potential
range and reach of threats threat phenomena attack surface threats to keep in view

Self-replicating Deepfakes and AI attack involves the Rapid machine-to-

AI-generated computational whole “AI stack”. machine escalation
malware propaganda Examples include: via automated C2

Improved Micro-targeting: Model inversion AI-enabled human

and autonomous AI-fused data for Training data augmentation by
disinformation targeting or blackmail manipulation peer competitors
campaigns “Data lake”
AI swarms and poisoning Proliferation of
AI-engineered and nano-swarms simple lethal
targeted pathogens autonomous weapons
to terrorists

Source: Final Report: National Security Commission on Artificial Intelligence

[email protected]

AI-driven malware, where a malicious logic embeds machine learning methods and models to
automatically: (i) probe the target system for inferring
actionable intelligence (e.g. system configuration or
operational patterns) and (ii) customize the attack
• First, digital dependence in all walks of life increases vulnerabilities to cyber intrusion
payload accordingly
across (e.g. determine
every segment the most
of our society: opportuneuniversities, government, private
corporations,
time toorganizations,
execute the payload
and the homes of individual impact).
so to maximize the citizens. In parallel, new sensors have
flooded the modern world. The internet of things (IoT), cars, phones, homes, and social
Attacksmedia platforms
Against collect streams
AI - Adversarial AI of data, which can then be fed into AI systems that can
identify, target, and manipulate
As AI proliferates, defeating adversariesor coerce
will be our citizens.on
predicated
2
defeating their AI and vice versa.
As Neural Networks
• Second, takenon-state
state and over sensor processing
adversaries areand triage tasks,
challenging the aUnited
human maybelow
States only be alerted
if the AIthe
deems it suspicious.
threshold Therefore,
of direct military we only need
confrontation to defeat
by using cyberthe AI to espionage,
attacks, evade detection, not
necessarily a human.and political warfare, and financial instruments. Adversaries do not
psychological
need AI to conduct widespread cyber attacks, exfiltrate troves of sensitive data about
American
Adversarial citizens,
attacks interfere
against in our
AI fall into elections,
three types:or bombard us with malign information on
• Data misclassification- to generate false positive or negative results
• Synthetic data generation-to feed false information
• Data analysis – for AI-assisted classical attack generation

AI Attack Surfaces
Electronic Attack (EA), Electronic Protection (EP), Electronic Support (ES) all have analogues in
the AI algorithmic domain. In the future, we may play the same game about the “Algorithmic
Spectrum,” denying our adversaries their AI capabilities while defending ours. Other can steal
or poison our models or manipulate our training data.
What Makes AI Possible Now?3

Four changes make Machine Learning possible now:

1. Massive Data Sets
2. Improved Machine Learning algorithms
3. Open-Source Code, Pretrained Models and
Frameworks
4. More computing power

Massive Data Sets

Machine Learning algorithms tend to require large quantities of training data in order to
produce high-performance AI models. (Training Google’s GPT-3 Natural Language Model with
175 billion parameters takes 1,024 Nvidia A100 GPUs more than one month.) Today, strategic
and tactical sensors pour in a firehose of images, signals and other data. Billions of computers,
digital devices and sensors connected to the Internet, producing and storing large volumes of
data, which provide other sources of intelligence. For example facial recognition requires
millions of labeled images of faces for training data.

Of course more data only helps if the data is relevant to your desired application. Training data
needs to match the real-world operational data very, very closely to train a high-performing AI
model.

Improved Machine Learning algorithms

The first Machine Learning algorithms are decades old, and some remain incredibly useful.
However, researchers have discovered new algorithms that have greatly sped up the fields
cutting-edge. These new algorithms have made Machine Learning models more flexible, more
robust, and more capable of solving different types of problems.

Open-Source Code, Pretrained Models and Frameworks

Developing Machine Learning systems required a lot of expertise and custom software
development that made it out of reach for most organizations. Now open-source code libraries
and developer tools allow organizations to use and build upon the work of external
communities. No team or organization has to start from scratch, and many parts that used to
require highly specialized expertise have been automated. Even non-experts and beginners can
create useful AI tools. In some cases, open-source ML models can be entirely reused and
purchased. Combined with standard competitions, open source, pretrained models and
frameworks have moved the field forward faster than any federal lab or contractor. It’s been a
feeding frenzy with the best and brightest researchers trying to one-up each other to prove
which ideas are best.

3
https://www.ai.mil/docs/Understanding%20AI%20Technology.pdf
The downside is that, unlike past DoD technology development - where the DoD leads it, can
control it, and has the most advanced technology (like stealth and electronic warfare), in most
cases the DoD will not have the most advanced algorithms or models. The analogy for AI is
closer to microelectronics than it is EW. The path forward for the DoD should be supporting
open research, but optimizing on data set collection, harvesting research results, and fast
application.

More computing power – special chips

Machine Learning systems require a lot of computing power. Today, it’s possible to run
Machine Learning algorithms on massive datasets using commodity Graphics Processing Units
(GPUs). (See the machine learning hardware section below). While many of the AI performance
improvements have been due to human cleverness on better models and algorithms, most of
the performance gains have been the massive increase in compute performance. (See the
semiconductor section.)

More computing power – AI In the Cloud

The rapid growth in the size of machine learning models has been achieved by the move to
large data center clusters. The size of machine learning models are limited by time to train
them. For example, in training images, the size of the model scales with the number of pixels in
an image. ImageNet Model sizes are 224x224 pixels. But HD (1920x1080) images require 40x
more computation/memory. Large Natural Language Processing models - e.g. summarizing
articles, English-to-Chinese translation like Google’s GPT-3 require enormous models. GPT-3
uses 175 billion parameters and was trained on a cluster with 1,024 Nvidia A100 GPUs that cost
~$25 million! (Which is why large clusters exist in the cloud, or the largest companies/
government agencies.) Facebooks Deep Learning and Recommendation Model (DLRM) was
trained on 1TB data and has 24 billion parameters. Some cloud vendors train on >10TB data
sets.

Instead of investing in massive amounts of computers needed for training companies can use
the enormous on-demand, off-premises hardware in the cloud (e.g. Amazon AWS, Microsoft
Azure) for both training machine learning models and deploying inferences.

We’re Just Getting Started

The next 10 years will see a massive improvement on AI inference and training capabilities. This
will require regular refreshes of the hardware– on the chip and cloud clusters - to take
advantage. This is the AI version of Moore’s Law on steroids – applications that are completely
infeasible today will be easy in 5 years.

What Can’t AI Do?

While AI can do a lot of things better than humans when focused on a narrow objective, there
are many things it still can’t do. AI works well in specific domain where you have lots of data,
time/resources to train, domain expertise to set the right goals/rewards during training, but
that is not always the case.
For example AI models are only as good as the fidelity and quality of the training data. Having
bad labels can wreak havoc on your training results. Protecting the integrity of the training data
is critical.

In addition, AI is easily fooled by out-of-domain data (things it hasn’t seen before). This can
happen by “overfitting” - when a model trains for too long on sample data or when the model is
too complex, it can start to learn the “noise,” or irrelevant information, within the dataset.4
When the model memorizes the noise and fits too closely to the training set, the model
becomes “overfitted,” and it is unable to generalize well to new data. If a model cannot
generalize well to new data, then it will not be able to perform the classification or prediction
tasks it was intended for. However, if you pause too early or exclude too many important
features, you may encounter the opposite problem, and instead, you may “underfit” your
model. Underfitting occurs when the model has not trained for enough time, or the input
variables are not significant enough to determine a meaningful relationship between the input
and output variables.

AI is also poor at estimating uncertainty /confidence (and explaining its decision-making). It

can’t choose its own goals. (Executives need to define the decision that the AI will execute.
Without well-defined decisions to be made, data scientists will waste time, energy and money.)
Except for simple cases an AI can’t (yet) figure out cause and effect or why something
happened. It can’t think creatively or apply common sense.

AI is not very good at creating a strategy (unless it can pull from previous examples and mimic
them, but then fails with the unexpected.) And it lacks generalized intelligence e.g. that can
generalize knowledge and translate learning across domains.

All of these are research topics actively being worked on. Solving these will take a combination
of high-performance computing, advanced AI/ML semiconductors, creative machine learning
implementations and decision science. Some may be solved in the next decade, at least to a
level where a human can’t tell the difference.

Where is AI in Business Going Next?

Skip this section if you’re interested in national security applications

Just as classic computers were applied to a broad set of business, science and military
applications, AI is doing the same. AI is exploding not only in research and infrastructure (which
go wide) but also in the application of AI to vertical problems (which go deep and depend more
than ever on expertise). Some of the new applications on the horizon include Human
AI/Teaming (AI helping in programming and decision making), smarter robotics and
autonomous vehicles, AI-driven drug discovery and design, healthcare diagnostics, chip
electronic design automation, and basic science research.

4
https://www.ibm.com/cloud/learn/overfitting
Advances in language understanding are being pursued to create systems that can summarize
complex inputs and engage through human-like conversation, a critical component of next-
generation teaming.

Where is AI and National Security Going Next?

In the near future AI may be able to predict the future actions an adversary could take and the
actions a friendly force could take to counter these. The 20th century model loop of Observe–
Orient–Decide and Act (OODA) is retrospective; an observation cannot be made until after the
event has occurred. An AI-enabled decision-making cycle might be ‘sense–predict–agree–act’:
AI senses the environment; predicts what the adversary might do and offers what a future
friendly force response should be; the human part of the human–machine team agrees with
this assessment; and AI acts by sending machine-to-machine instructions to the small, agile and
many autonomous warfighting assets deployed en masse across the battlefield.

An example of this is DARPA’s ACE (Air Combat Evolution) program that is developing a
warfighting concept for combined arms using a manned and unmanned systems. Humans will
fight in close collaboration with autonomous weapon systems in complex environments with
tactics informed by artificial intelligence.

A Once-in-a-Generation Event
Imagine it’s the 1980’s and you’re in charge of an intelligence agency. SIGINT and COMINT were
analog and RF. You had worldwide collection systems with bespoke systems in space, air,
underwater, etc. And you wake up to a world that shifts from copper to fiber. Most of your
people, and equipment and equipment are going to be obsolete, and you need to learn how to
capture those new bits. Almost every business processes needed to change, new organizations
needed to be created, new skills were needed, and old ones were obsoleted. That’s what AI/ML
is going to do to you and your agency.

The primary obstacle to innovation in national security is not technology, it is culture. The DoD
and IC must overcome a host of institutional, bureaucratic, and policy challenges to adopting
and integrating these new technologies. Many parts of our culture are resistant to change,
reliant on traditional tradecraft and means of collection, and averse to risk-taking, (particularly
acquiring and adopting new technologies and integrating outside information sources.)

History tells us that late adopters fall by the wayside as more agile and opportunistic
governments master new technologies.

Carpe Diem.

Want more Detail?

Read on if you want to know about Machine Learning chips, see a sample Machine Learning
Pipeline and learning about the four types of Machine Learning.
Artificial Intelligence/Machine Learning Semiconductors
Skip this section if all you need to know is that special chips are used for AI/ML.

AI/ML, semiconductors, and high-performance computing are intimately intertwined - and

progress in each is dependent on the others. (See the “Semiconductor Ecosystem” report.)

Some machine learning models can have trillions of parameters and require a massive number
of specialized AI chips to run. Edge computers are significantly less powerful than the massive
compute power that’s located at data centers and the cloud. They need low power and
specialized silicon.

Why Dedicated AI Chips and Chip Speed Matter

Dedicated chips for neutral nets (e.g. Nvidia GPUs, Xilinx FPUs, Google TPUs) are faster than
conventional CPUs for three reasons: 1) they use parallelization, 2) they
have larger memory bandwidth and 3) they have fast memory access.

There are three types of AI Chips:

• Graphics Processing Units (GPUs) - Thousands of cores, parallel
workloads, widespread use in machine learning
• Field-Programmable Gate Arrays (FPGAs) - Good for algorithms;
compression, video encoding, cryptocurrency, genomics, search.
Needs specialists to program,
• Application-Specific Integrated Circuits (ASICs) – custom chips e.g.
Google TPU’s

Matrix multiplication plays a big part in neural network computations, especially if there are
many layers and nodes. Graphics Processing Units (GPUs) contain 100s or 1,000s of cores that
can do these multiplications simultaneously. And neural networks are inherently parallel which
means that it’s easy to run a program across the cores and clusters of these processors. That
makes AI chips 10s or even 1,000s of times faster and more efficient than classic CPUs for
training and inference of AI algorithms. State-of-the-art AI chips are dramatically more cost-
effective than state-of-the-art CPUs as a result of their greater efficiency for AI algorithms.

Cutting-edge AI systems require not only AI-specific chips, but state-of-the-art AI chips. Older AI
chips incur huge energy consumption costs that quickly balloon to unaffordable levels. Using
older AI chips today means overall costs and slowdowns at least an order of magnitude greater
than for state-of- the-art AI chips.
Cost and speed make it virtually impossible to develop and deploy cutting-edge AI algorithms
without state-of-the-art AI chips. Even with state-of-the-art AI chips, training a large AI
algorithm can cost tens of millions of dollars and take weeks to
complete. With general-purpose chips like CPUs or older AI chips,
this training would take much longer and cost orders of
magnitude more, making staying at the R&D frontier impossible.
Similarly, performing inference using less advanced or less
specialized chips could involve similar cost overruns and take
orders of magnitude longer.

In addition to off-the-shelf AI chips from Nvidia, Xlinix and Intel, large companies like Facebook,
Google, Amazon, have designed their own chips to accelerate AI. The opportunity is so large
that there are hundreds of AI accelerator startups designing their own chips, funded by 10’s of
billions of venture capital and private equity. None of these companies own a chip
manufacturing plant (a fab) so they all use a foundry (an independent company that makes
chips for others) like TSMC in Taiwan (or SMIC in China for Defense related silicon.)

A Sample of AI GPU, FPGA and ASIC AI Chips and Where They’re Made

IP (Intellectual Property) Vendors Also Offer AI Accelerators

AI chip designers can buy AI IP Cores – prebuilt AI accelerators from Synopsys (EV7x,) Cadence
(Tensilica AI,) Arm (Ethos,) Ceva (SensPro2, NeuPro), Imagination (Series4,) ThinkSilicon (Neox,)
FlexLogic (eFPGA,) Edgecortix and others.

Other AI Hardware Architectures

Spiking Neural Networks (SNN) is a completely different approach from Deep Neural Nets. A
form of Neuromorphic computing it tries to emulate how a brain works. SNN neurons use
simple counters and adder—no matrix multiply hardware is needed and power consumption is
much lower. SNNs are good at unsupervised learning – e.g. detecting patterns in unlabeled data
streams. Combined with their low power they’re a good fit for sensors at the edge. Examples::
BrainChip, GrAI Matter, Innatera, Intel.

Analog Machine Learning AI chips use analog circuits to do the matrix multiplication in memory.
The result is extremely low power AI for always-on sensors. Examples: Mythic (AMP,) Aspinity
(AML100,) Tetramem.

Optical (Photonics) AI Computation promise performance gains over standard digital silicon,
and some are nearing production. They use intersecting coherent light beams rather than
switching transistors to perform matrix multiplies. Computation happens in picoseconds and
requires only power for the laser. (Though off-chip digital transitions still limit power savings.)
Examples: Lightmatter, Lightelligence, Luminous, Lighton.

AI Hardware for the Edge

As more AI moves to the edge, the Edge AI accelerator market is segmenting into high-end
chips for camera-based systems and low-power chips for simple sensors. For example:

AI Chips in Autonomous vehicles, Augmented Reality and multicamera surveillance systems

These inference engines require high performance. Examples: Nvidia (Orin,) AMD (Versal,)
Qualcomm (Cloud AI 100,) and acquired Arriver for automotive software.

AI Chips in Cameras for facial recognition, surveillance. These inference chips require a balance
of processing power with low power. Putting an AI chip in each camera reduces latency and
bandwidth. Examples: Hailo-8, Ambarella CV5S, Quadric (Q16), (RealTek 3916N).
Ultralow-Power AI Chips Target IoT Sensors - IoT devices require very simple neural networks
and can run for years on a single battery. Example applications: Presence detection, wakeword
detection, gunshot detection... Examples: Syntiant (NDP,) Innatera, BrainChip,,

AI/ML Hardware Benchmarks

While there are lots of claims about how much faster each of these chips are for AI/ML there
are now a set of standard benchmarks - MLCommons. These benchmarks were created by
Google, Baidu, Stanford, Harvard and U.C. Berkeley.

One Last Thing - Non-Nvidia AI Chips and the “Nvidia Software Moat”
New AI accelerator chips most deal with the software moat that Nvidia has built around their
GPU’s. As popular AI applications and frameworks are built on Nvidia CUDA software platform,
if new AI Accelerator vendors want to port these applications to their chips they have to build
their own drivers, compiler, debugger, and other tools.
Details of a machine learning pipeline
This is a sample of the workflow (a pipeline) data scientists use to develop, deploy and maintain
a machine learning model (see the detailed description here.)

The Types of Machine Learning5 – skip this section if you want to believe it’s
magic.

Machine Learning algorithms fall into four classes:

1. Supervised Learning
2. Unsupervised Learning
3. Semi-supervised Learning
4. Reinforcement Learning

They differ based on:

• What types of data their algorithms can work with
• For supervised and unsupervised learning, whether or not the training data is labeled or
unlabeled
• How the system receives its data inputs

5
https://www.ai.mil/docs/Understanding%20AI%20Technology.pdf
 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.
Supervised Learning The four categories – Figure 3. Labeled and Unlabeled Training Data
• A “supervisor” (a human or a software explained
system)more accurately
the following page –
on
labels each of the training data
inputs with its correct associated output
differ based on what
types of data their
• Note that pre-labeled data is only required forcan
algorithms thework
with. However, the
training data that
the algorithm uses to train the AI modeimportant distinction
is not whether the
• In operation in the inference phase thedataAI willis beaudio,
generating its own
images, text, or
labels, the accuracy of which will depend on the AI’s
numbers. Rather, the training
important distinction
• Supervised Learning can achieve extremely high
is whether or notperformance, but they
the training data is labeled require
or unlabeled very
and how the large,
system
receives its data inputs. Figure 3 provides a simple illustration of labeled and
labeled datasets unlabeled training data for a classifier of images of cats and dogs.

• Using labeled inputs and outputs, the model

Depending can
uponmeasure
whether or not itsdata
accuracy
is labeled, aand learn
different family over time
of algorithms
applies. The four major families of algorithms are Supervised Learning,
• For images a rule of thumb is that the algorithm needs at least 5,000 labeled examples
Unsupervised Learning, Semi-Supervised Learning, and Reinforcement Learning. of
each category in order to produce an AISupervised
modelLearning:
with decent
“Supervised” performance
means that, before the algorithm processes

• In supervised learning, the algorithm “learns” from

or a different the
software training
system) dataset
has accurately byeach
the training data, some “supervisor” (which may be a human, group of humans,
labeled iteratively
of the data inputs
making predictions on the data and adjusting for the correct answer. 11
Gregory C. Allen | DoD Joint AI Center
• While supervised learning models tend to be more accurate than unsupervised learning
models, they require upfront human intervention to label the data appropriately.

Supervised Machine Learning - Categories and Examples:

• Classification problems - use an algorithm to assign data into specific categories, such as
separating apples from oranges. Or classify spam in a separate folder from your inbox.
Linear classifiers, support vector machines, decision trees and random forest are all
common types of classification algorithms.
• Regression - understands the relationship between dependent and independent variables.
Helpful for predicting numerical values based on different data points, such as sales
revenue projections for a given business. Some popular regression algorithms are linear
regression, logistic regression and polynomial regression.
• Example algorithms include: Logistic Regression and the Back Propagation Neural Network

Unsupervised Learning
• These algorithms can analyze and cluster unlabeled data sets. They discover hidden
patterns in data without the need for human intervention (hence, they are “unsupervised”)
• They can extract features from the data without a label for the results
• For an image classifier, an unsupervised algorithm would not identify the image as a “cat”
or a “dog.” Instead, it would sort the training dataset into various groups based on their
similarity
• Unsupervised Learning systems are often less predictable, but as unlabeled data is usually
more available than labeled data, they are important
• Unsupervised algorithms are useful when developers want to understand their own
datasets and see what properties might be useful in either developing automation or
change operational practices and policies
• They still require some human intervention for validating the output
Unsupervised Machine Learning - Categories and Examples Understanding Artificial Intelligence Technology 13
• Clustering groups unlabeled data based on their similarities or differences. For example, K-
means clustering algorithms assign similar data pointsusing
into
Figure groups,
4 depicts where
the differences
an image analysis example.
betweenthe
SupervisedK value
and Unsupervised algorithms

represents the size of the grouping and granularity. This technique is helpful for market
Figure 4. Illustrated Example of Supervised and Unsupervised Algorithms

segmentation, image compression, etc.

• Association finds relationships between variables in a given dataset. These
methods are frequently used for market basket analysis and
recommendation engines, along the lines of “Customers Who Bought This
Item Also Bought” recommendations.
• Dimensionality reduction is used when the number of features (or
dimensions) in a given dataset is too high. It reduces the number of data
inputs to a manageable size while also preserving the data integrity. Often,
this technique is used in the preprocessing data stage, such as when
autoencoders remove noise from visual data to improve picture quality.
• Example algorithms include: Apriori algorithm and K-Means

Difference between supervised and unsupervised learning

The main difference: Labeled data It is not true that Unsupervised Learning is “worse” than Supervised Learning
(though performance can be lower for some use cases). Rather, Unsupervised
• Goals: In supervised learning, the goal is to predict outcomes
use case is fraud for newin financial
detection data.data. YouIn theknow
case ofup
Learning is useful for solving different types of problems. A common Unsupervised
fraud detection,

front the type of results to expect. With an unsupervised

matcheslearning algorithm, the goal is to
Supervised Learning could be a good fit for identifying potential fraud that
behaviors known to be unlawful or associated with fraud. Unsupervised
Learning can find new, unidentified patterns of behavior that might indicate new
get insights from large volumes of new data. The machine learning itself determines what is
types of fraud techniques.

different or interesting from the dataset.

Gregory C. Allen | DoD Joint AI Center 13
• Applications: Supervised learning models are ideal for spam detection, sentiment analysis,
weather forecasting and pricing predictions, among other things. In contrast, unsupervised
learning is a great fit for anomaly detection, recommendation engines, customer personas
and medical imaging.
• Complexity: Supervised learning is a simple method for machine learning, typically
calculated through the use of programs like R or Python. In unsupervised learning, you need
powerful tools for working with large amounts of unclassified data. Unsupervised learning
models are computationally complex because they need a large training set to produce
intended outcomes.
• Drawbacks: Supervised learning models can be time-consuming to train, and the labels for
input and output variables require expertise. Meanwhile, unsupervised learning methods
can have wildly inaccurate results unless you have human intervention to validate the
output variables.

Semi-Supervised Learning
• “Semi- Supervised” algorithms 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
 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

• Typical algorithms are extensions to other flexible methods that make assumptions about
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
how to model the unlabeled data. An example is Unsupervised Learning:

Generative Adversarial Networks trained on 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

photographs can generate new photographs that look 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
authentic to human observers (deep fakes) 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

Reinforcement Learning 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
• Training data is collected by an autonomous, self-directed AI agent as it perceives its
receipt of cumulative rewards.

environment and performs goal-directed actions Figure 5: Simplified Reinforcement Learning Diagram

• The rewards are input data received by the AI agent

when certain criteria are satisfied.
• These criteria are typically unknown to the agent at
the start of training
• Rewards often contain only partial information.
They don’t signal which inputs were good or not
• The system is learning to take actions to maximize its receipt of cumulative rewards
Gregory C. Allen | DoD Joint AI Center 14

• Reinforcement AI can defeat humans– in chess, Go…

• There are no labeled datasets for every possible move
• There is no assessment of whether it was a “good or bad move
• Instead, partial labels reveal the final outcome “win” or “lose”
• The algorithms explore the space of possible actions to learn the optimal set of rules for
determining the best action that maximize wins

Reinforcement Machine Learning - Categories and Examples

• Algorithm examples include: DQN (Deep Q Network), DDPG (Deep Deterministic Policy
Gradient), A3C (Asynchronous Advantage Actor-Critic Algorithm), NAF (Q-Learning with
Normalized Advantage Functions), …
• AlphaGo, a Reinforcement system played 4.9 million
games of Go in 3 days against itself to learn how to
play the game at a world-champion level
• Reinforcement is challenging to use in the real
world, as the real world is not as heavily bounded as
video games and time cannot be sped up in the real
world
• There are consequences to failure in the real world

Sources:
• Understanding AI Technology: Greg Allen, Chief of Strategy and Communications Joint
Artificial Intelligence Center (JAIC), Department of Defense
https://www.ai.mil/docs/Understanding%20AI%20Technology.pdf
• AI, Machine Learning, Deep Learning Explained Simply: Jun Wu
https://towardsdatascience.com/ai-machine-learning-deep-learning-explained-simply-
7b553da5b960
• The Democratization of Artificial Intelligence and Deep Learning: Databricks
https://databricks.com/discover/pages/the-democratization-of-artificial-intelligence-and-
deep-learning
• Final Report: National Security Report on Artificial Intelligence https://www.nscai.gov/wp-
content/uploads/2021/03/Full-Report-Digital-1.pdf
• A Beginners Guide to Neural Nets and Deep Learning: Pathmind
https://wiki.pathmind.com/neural-network