Expansion of wargames and military case studies that introduce challenging dilemmas at the intersection of technology and military art will go a long way towards cultivating the mindsets needed in the 21st century service member.
Troops on the front lines should be the leaders of driving innovations in this space. Recently released planning guidance by the Commandant of the Marine Corps recognizes the importance of individuals with artificial intelligence skill sets.
Developing programs that mimic human behaviors such as reasoning or judgement is a non-trivial matter. Hard-coding general human-like behavior is also unrealistic, because the code required would quickly become unwieldy in accounting for countless combinations of events or changes in contextual factors. Artificial intelligence technologies overcome the shortcomings of hard-coding human-like behaviors through a variety of techniques, but they also trade increased performance for greater opacity.
For instance, under the umbrella of machine learning, techniques such as deep learning e.g., neural networks, convolutional neural networks, etc. use feature extraction techniques composed of many neural layers fed millions of examples to train and adjust the algorithm’s parameters. Thus, natural language processing and self-driving cars are made possible, not through hard-coding explicit instructions, but through different types and combinations of deep learning methods.
Understanding the results of algorithms that ingest petabytes of data, however, presents a black-box problem for consumers of such system outputs. The march towards greater automation will further conflate the lines between human and machine decision-making where a greater number of outputs are inextricably tied to artificial intelligence and machine learning technologies.
In an AI Demo, robot was used to demonstrate how well the technology allowed robots to understand instructions. Two of the three demonstrations went off perfectly. The robot had to be rebooted during the third when its navigation system locked up.
Trust is the key to getting robots and humans to work together. Soldiers will need to learn the robot’s capabilities and limitations, and at the same time, the machine will learn the unit’s language and procedures.
Big challenges remain. First, many robot are currently too slow for practical use. Second, it needs to be far more resilient. All AI systems can go wrong, but military robots have to be reliable in life-and-death situations. These challenges are being addressed.
The big questions for many with the task of designing robots for practical use involve achieving speed and resiliency: How will it do so, when it will do so, and in which markets and applications will it have the most impact? Certainly, military field-level computing applications and virtual work spaces are among those most clearly in the crosshairs of machine learning and professional spaces.
Artificial intelligence tech is limited by challenge of accepting outputs from artificial intelligence tools, especially those outputs based on opaque processes. It’s one reason why policy makers were historically slow to embrace artificial intelligence, and particularly neural networks, for decision making.
Neural nets and deep learning approaches have proven effective in chaotic, unstructured field scenarios and problems, such as helping self-driving cars navigate city streets, helping algorithms identify objects in YouTube videos, or helping computer games win the high-variable game of Go. But the underlying processes behind them defy easy explanation, and that’s a big drawback.
“The biggest part of the problem of artificial intelligence is: they build these incredibly long algorithms with all of these gates to go through. They push all of this machine learning and data through it. Frankly, we are not entirely sure how all of that works, all the time.
“If someone sabotages your data that you are feeding your algorithm for learning, we’re not entirely sure what’s going to come out of the other end. It’s going to take a while to figure out how to organize these things to be successful.
Wouldn’t it be great if we could shoot someone in the face at 200 kilometers? They don’t even know you are there. That’s the kind of man-machine teaming we really want to get after, but Commanders need some way to better explain why they arrived at the decision they did, besides “the machine made me do it.”
“We haven’t cracked the nut on man-machine teaming yet. No one really has. The closest we’ve gotten is the extremely high level of information we’ve pushed to aviators in cockpits.
Pattern-recognition capabilities of AI with potential to enable applications ranging from machines that spot and zap bugs to apps that, with the help of augmented-reality displays on smart phones, enable troops to diagnose and solve problems with mission critical gear.
AI programs are striving to clearly explain the basis of their actions and how they arrive at particular decisions. AI that can explain itself should enable users to trust it, a good thing as increasingly complex AI-driven systems become commonplace.
Early AI researchers focused on chess because it presents a difficult intellectual challenge for humans, yet the rules are simple enough to describe easily in a computer programming language.
Chess champions use knowledge of the game to ignore most potential moves that would make no sense to execute. The first AI chess programs used heuristics, or rules of thumb, to decide which moves to spend time considering.
Decades ago, computers were automating boring and laborious tasks, like payroll accounting, or solving complex mathematical equations, such as plotting the trajectories of the Apollo missions to the moon.
Not surprisingly, AI researchers ignored the boring applications of computers and instead conceived of artificial intelligence as computers solving complex mathematical equations, expressed as algorithms.
Algorithms are sets of simple instructions that computers execute in sequence to produce results, such as calculating the trajectory of a lunar lander, when it should fire its retro rockets, and for how long.
As the centrality of knowledge to intelligence became apparent, AI researchers focused on building so-called expert systems. These programs captured the specialized knowledge of experts in rules that they could then apply to situations of interest to generate useful results.
If you’ve ever used a program such as TurboTax to prepare your income tax return, you’ve used an expert system. It became apparent that expert systems were difficult to update and maintain, and they would give bizarrely wrong answers when confronted with unusual inputs.
The hype around AI gave way to disappointment and the term AI fell out of favor and was superseded by terms such as distributed agents, probabilistic reasoning, and neural networks.
Later, another approach to AI was gaining momentum. Rather than focus on explicitly writing down knowledge, why not try to create machines that learn the way people do? A robot that could learn from people, observations, and experience should be able to get around in the world, stopping to ask for directions or calling for help when necessary.
So-called machine-learning approaches try to extract useful knowledge directly from data about the world. Rather than structuring this knowledge as rules, machine-learning systems apply statistical and probabilistic methods to create generalizations from many data points. The resulting systems are not always correct, but then again, neither are people. Being right most of the time is sufficient for many real-world tasks.
Artificial neural networks that mimic capabilities of the brain in order to recognize patterns. Instead of writing computer code to program these networks, researchers train them. A common method is called supervised learning, in which researchers collect a large set of data, such as photos of objects to be recognized, and label them appropriately.
Artificial intelligence has developed in two major waves. The first wave focused on hand-crafted knowledge, in which experts characterized their understanding of a particular area, such as income tax return preparation, as a set of rules. The second wave focused on machine-learning, which creates pattern-recognition systems by training on large sets of data. The resulting systems are surprisingly good at recognizing objects, such as faces.
DARPA believes that the next major wave of progress will combine techniques from the first and second waves to create systems that can explain their outputs and apply commonsense reasoning to act as problem-solving partners.
Deep neural networks can make use of more data to improve their recognition accuracy well past the point at which other approaches cease improving. This superior performance has made deep networks the mainstay of the current wave of AI applications.
By extracting knowledge directly from data, neural networks avoid the need to write down rules that describe the world. This approach makes them better for capturing knowledge that’s hard to describe in words.
Autonomous Land Vehicle program developed self-driving cars. Both teams used neural networks to enable their vehicles to recognize the edges of the road. However, the systems were easily confused by leaves or muddy tire tracks on the road, because the hardware available at the time was not powerful enough. Nonetheless, the program established the scientific and engineering foundations of autonomous vehicles.
But deep neural networks are very inefficient learners, requiring millions of images to learn how to detect objects. They are better thought of as statistical pattern recognizers produced by an algorithm that maps the contours of the training data. Give these algorithms enough pictures of objects, and they will find the differences that distinguish the one from the other. For some applications, this inefficiency is not an issue.
Internet search engines can now find pictures of just about anything. For applications where training data is scarce, neural networks can generate it. An approach called generative adversarial networks takes a training set of pictures and pits Digital Twin networks against each other.
For application of Digital Twin Tool, one tries to generate new pictures that are similar to the training set, and the other tries to detect the generated pictures. Over multiple rounds, the two networks get better at generation and detection, until the pictures produced are novel, yet usefully close to real ones, so that they can be used to augment a training set. Note that no labels are required for this generation phase, as the objective is to generate new pictures, not classify existing ones.
DARPA is running a program to develop systems that can produce accurate explanations at the right level for a user. Systems that can explain themselves will enable more effective human/machine partnerships.
Another approach to machine learning relies on cues from the environment to reinforce good behavior and suppress bad behavior. For example, the program AlphaGo Zero can teach itself to play the board game Go at championship levels without any human input, other than the rules of the game. It starts by playing against itself, making random moves. It uses the rules of the game to score its results, and these scores reinforce winning tactics.
This so-called reinforcement learning can be highly effective in situations where there are clear rewards for effective behavior. However, determining what behavior created the desired result in many real-world situations can be difficult.
As AI increasingly makes its way into industrial settings and consumer products, companies are discovering that its substantial benefits come with costs, in the form of engineering complexity and unique requirements for ongoing maintenance.
The computational intensity of AI systems requires racks of servers and networking gear, which must be secured against and continuously monitored for intrusions. The unlimited appetite of these systems for data often makes them dependent on many different enterprise databases, which requires ever-increasing coordination of operations across the organization.
Machine-learning systems must be continually retrained to keep them in sync with the world as it continually changes and evolves. Ultimately, people are still far more effective learners than machines. We can learn from teachers, books, observation, and experience. We can quickly apply what we’ve learned to new situations, and we learn constantly in daily life. We can also explain our actions, which can be quite helpful during the learning process.
In contrast, deep learning systems do all their learning in a training phase, which must be complete before they can reliably recognize things in the world. Trying to learn while doing can create catastrophic forgetting, as backpropagation makes wholesale changes to the link weights between the nodes of the neural network.
DARPA Learning Machines program is exploring ways to enable machines to learn while doing without catastrophic forgetting. Such a capability would enable systems to improve on the fly, recover from surprises, and keep them from drifting out of sync with the world.
The real breakthrough for artificial intelligence will come when researchers figure out a way to learn or otherwise acquire common sense. Without common sense, AI systems will be powerful but limited tools that require human inputs to function. With common sense, an AI could become a partner in problem-solving.
The knowledge of a trained neural network is contained in the thousands of weights on its links. This encoding prevents neural networks from explaining their results in any meaningful way. DARPA is currently running a program called Explainable AI to develop new machine-learning architectures that can produce accurate explanations of their decisions in a form that makes sense to humans. As AI algorithms become more widely used, reasonable self-explanation will help users understand how these systems work, and how much to trust them in various situations.
Once trained, current machine-learning systems no longer adapt to their environments. DARPA’s Lifelong Learning Machines program is researching ways to enable systems to learn from surprises and adapt to changes in their environments. The Assured Autonomy program is developing approaches to produce mathematical assurance that such systems will operate safety and predictably under a wide range of operating conditions.
In combination with large data sets and libraries, improvements in computer performance over the last decade have enabled the success of machine learning. More performance at lower electrical power is essential to allow this use of AI for data-center applications and for tactical deployments.
DARPA has demonstrated analog processing of AI algorithms that operate a thousand times faster using a thousand times less power compared to state-of-the-art digital processors. New research will investigate AI-specific hardware designs and address the inefficiency of machine learning by drastically reducing requirements for labeled training data.
DARPA has taken the lead in pioneering research to develop the next generation of AI algorithms, which will transform computers from tools into problem-solving partners. New research will enable AI systems to acquire and reason with commonsense knowledge. DARPA R&D produced the first AI successes, such as expert systems and search utilities, and more recently has advanced machine-learning tools and hardware.
Current AI systems today seem superhuman because they can do complex reasoning quickly in narrow specialties, but creates an illusion that they are smarter and more capable than they really are.
Roadmap will build upon these strategies by listing the types of AI that will be needed year-to-year to support military strategy and maintain a firm understanding of what AI is and how it will be used to benefit the organisation. This understanding should go beyond buzzwords and definitions.
Compromise of data through process models could mean the compromise of sources and methods. When AI is introduced, new attack vectors are introduced, such as deep learning spoofing and data spoofing.
If the training data is known or manipulation of data is too predictable, adversaries can easily anticipate and predict actions and outcomes.
Adversaries can spoof sensors and the data collected by those sensors without needing to mess with the underlying model code. Put simply, adversaries do not need to know what is in the box to exploit the box.
1. Limitation of Data utilisation/Availability and bias of programmers and embedded in data sets
2. Shortage of strategic approach to understand implementation times and integration challenges
3. Not enough Usablity/Interoperablity with other systems/platforms and ability to decipher how AI arrives at decisions
4. Trouble transferring learning from one experience to another, and does not understand causal reasoning
5. Lack of explanation capability and unable to do complex future planning
6. Difficulty handling unexpected circumstances, trouble dealing with boundary conditions and lack of context dependent learning
7. Can’t decide it's own learning algorithm based on situation or carry out self planning about the best topology structure to use
8. Questions remain about ability to demonstrate multi-domain integrated learning
9. Not enough computing power
10. Vulnerable to adversary attack