Tag Archives: how

#439875 Not So Mysterious After All: Researchers ...

The deep learning neural networks at the heart of modern artificial intelligence are often described as “black boxes” whose inner workings are inscrutable. But new research calls that idea into question, with significant implications for privacy.

Unlike traditional software whose functions are predetermined by a developer, neural networks learn how to process or analyze data by training on examples. They do this by continually adjusting the strength of the links between their many neurons.

By the end of this process, the way they make decisions is tied up in a tangled network of connections that can be impossible to follow. As a result, it’s often assumed that even if you have access to the model itself, it’s more or less impossible to work out the data that the system was trained on.

But a pair of recent papers have brought this assumption into question, according to MIT Technology Review, by showing that two very different techniques can be used to identify the data a model was trained on. This could have serious implications for AI systems trained on sensitive information like health records or financial data.

The first approach takes aim at generative adversarial networks (GANs), the AI systems behind deepfake images. These systems are increasingly being used to create synthetic faces that are supposedly completely unrelated to real people.

But researchers from the University of Caen Normandy in France showed that they could easily link generated faces from a popular model to real people whose data had been used to train the GAN. They did this by getting a second facial recognition model to compare the generated faces against training samples to spot if they shared the same identity.

The images aren’t an exact match, as the GAN has modified them, but the researchers found multiple examples where generated faces were clearly linked to images in the training data. In a paper describing the research, they point out that in many cases the generated face is simply the original face in a different pose.

While the approach is specific to face-generation GANs, the researchers point out that similar ideas could be applied to things like biometric data or medical images. Another, more general approach to reverse engineering neural nets could do that straight off the bat, though.

A group from Nvidia has shown that they can infer the data the model was trained on without even seeing any examples of the trained data. They used an approach called model inversion, which effectively runs the neural net in reverse. This technique is often used to analyze neural networks, but using it to recover the input data had only been achieved on simple networks under very specific sets of assumptions.

In a recent paper, the researchers described how they were able to scale the approach to large networks by splitting the problem up and carrying out inversions on each of the networks’ layers separately. With this approach, they were able to recreate training data images using nothing but the models themselves.

While carrying out either attack is a complex process that requires intimate access to the model in question, both highlight the fact that AIs may not be the black boxes we thought they were, and determined attackers could extract potentially sensitive information from them.

Given that it’s becoming increasingly easy to reverse engineer someone else’s model using your own AI, the requirement to have access to the neural network isn’t even that big of a barrier.

The problem isn’t restricted to image-based algorithms. Last year, researchers from a consortium of tech companies and universities showed that they could extract news headlines, JavaScript code, and personally identifiable information from the large language model GPT-2.

These issues are only going to become more pressing as AI systems push their way into sensitive areas like health, finance, and defense. There are some solutions on the horizon, such as differential privacy, where models are trained on the statistical features of aggregated data rather than individual data points, or homomorphic encryption, an emerging paradigm that makes it possible to compute directly on encrypted data.

But these approaches are still a long way from being standard practice, so for the time being, entrusting your data to the black box of AI may not be as safe as you think.

Image Credit: Connect world / Shutterstock.com Continue reading

Posted in Human Robots

#439820 How Musicologists and Scientists Used AI ...

When Ludwig van Beethoven died in 1827, he was three years removed from the completion of his Ninth Symphony, a work heralded by many as his magnum opus. He had started work on his 10th Symphony but, due to deteriorating health, wasn’t able to make much headway: All he left behind were some musical sketches.

Ever since then, Beethoven fans and musicologists have puzzled and lamented over what could have been. His notes teased at some magnificent reward, albeit one that seemed forever out of reach.

Now, thanks to the work of a team of music historians, musicologists, composers and computer scientists, Beethoven’s vision will come to life.

I presided over the artificial intelligence side of the project, leading a group of scientists at the creative AI startup Playform AI that taught a machine both Beethoven’s entire body of work and his creative process.

A full recording of Beethoven’s 10th Symphony is set to be released on Oct. 9, 2021, the same day as the world premiere performance scheduled to take place in Bonn, Germany—the culmination of a two-year-plus effort.

Past Attempts Hit a Wall
Around 1817, the Royal Philharmonic Society in London commissioned Beethoven to write his ninth and 10th symphonies. Written for an orchestra, symphonies often contain four movements: the first is performed at a fast tempo, the second at a slower one, the third at a medium or fast tempo, and the last at a fast tempo.

Beethoven completed his Ninth Symphony in 1824, which concludes with the timeless “Ode to Joy.”

But when it came to the 10th Symphony, Beethoven didn’t leave much behind, other than some musical notes and a handful of ideas he had jotted down.

A page of Beethoven’s notes for his planned 10th Symphony. Image Credit: Beethoven House Museum, CC BY-SA

There have been some past attempts to reconstruct parts of Beethoven’s 10th Symphony. Most famously, in 1988, musicologist Barry Cooper ventured to complete the first and second movements. He wove together 250 bars of music from the sketches to create what was, in his view, a production of the first movement that was faithful to Beethoven’s vision.

Yet the sparseness of Beethoven’s sketches made it impossible for symphony experts to go beyond that first movement.

Assembling the Team
In early 2019, Dr. Matthias Röder, the director of the Karajan Institute, an organization in Salzburg, Austria, that promotes music technology, contacted me. He explained that he was putting together a team to complete Beethoven’s 10th Symphony in celebration of the composer’s 250th birthday. Aware of my work on AI-generated art, he wanted to know if AI would be able to help fill in the blanks left by Beethoven.

The challenge seemed daunting. To pull it off, AI would need to do something it had never done before. But I said I would give it a shot.

Röder then compiled a team that included Austrian composer Walter Werzowa. Famous for writing Intel’s signature bong jingle, Werzowa was tasked with putting together a new kind of composition that would integrate what Beethoven left behind with what the AI would generate. Mark Gotham, a computational music expert, led the effort to transcribe Beethoven’s sketches and process his entire body of work so the AI could be properly trained.

The team also included Robert Levin, a musicologist at Harvard University who also happens to be an incredible pianist. Levin had previously finished a number of incomplete 18th-century works by Mozart and Johann Sebastian Bach.

The Project Takes Shape
In June 2019, the group gathered for a two-day workshop at Harvard’s music library. In a large room with a piano, a blackboard and a stack of Beethoven’s sketchbooks spanning most of his known works, we talked about how fragments could be turned into a complete piece of music and how AI could help solve this puzzle, while still remaining faithful to Beethoven’s process and vision.

The music experts in the room were eager to learn more about the sort of music AI had created in the past. I told them how AI had successfully generated music in the style of Bach. However, this was only a harmonization of an inputted melody that sounded like Bach. It didn’t come close to what we needed to do: construct an entire symphony from a handful of phrases.

Meanwhile, the scientists in the room—myself included—wanted to learn about what sort of materials were available, and how the experts envisioned using them to complete the symphony.

The task at hand eventually crystallized. We would need to use notes and completed compositions from Beethoven’s entire body of work—along with the available sketches from the 10th Symphony—to create something that Beethoven himself might have written.

This was a tremendous challenge. We didn’t have a machine that we could feed sketches to, push a button and have it spit out a symphony. Most AI available at the time couldn’t continue an uncompleted piece of music beyond a few additional seconds.

We would need to push the boundaries of what creative AI could do by teaching the machine Beethoven’s creative process—how he would take a few bars of music and painstakingly develop them into stirring symphonies, quartets, and sonatas.

Piecing Together Beethoven’s Creative Process
As the project progressed, the human side and the machine side of the collaboration evolved. Werzowa, Gotham, Levin, and Röder deciphered and transcribed the sketches from the 10th Symphony, trying to understand Beethoven’s intentions. Using his completed symphonies as a template, they attempted to piece together the puzzle of where the fragments of sketches should go—which movement, which part of the movement.

They had to make decisions, like determining whether a sketch indicated the starting point of a scherzo, which is a very lively part of the symphony, typically in the third movement. Or they might determine that a line of music was likely the basis of a fugue, which is a melody created by interweaving parts that all echo a central theme.

The AI side of the project—my side—found itself grappling with a range of challenging tasks.

First, and most fundamentally, we needed to figure out how to take a short phrase, or even just a motif, and use it to develop a longer, more complicated musical structure, just as Beethoven would have done. For example, the machine had to learn how Beethoven constructed the Fifth Symphony out of a basic four-note motif.

Next, because the continuation of a phrase also needs to follow a certain musical form, whether it’s a scherzo, trio, or fugue, the AI needed to learn Beethoven’s process for developing these forms.

The to-do list grew: We had to teach the AI how to take a melodic line and harmonize it. The AI needed to learn how to bridge two sections of music together. And we realized the AI had to be able to compose a coda, which is a segment that brings a section of a piece of music to its conclusion.

Finally, once we had a full composition, the AI was going to have to figure out how to orchestrate it, which involves assigning different instruments for different parts.
And it had to pull off these tasks in the way Beethoven might do so.

Passing the First Big Test
In November 2019, the team met in person again—this time, in Bonn, at the Beethoven House Museum, where the composer was born and raised.

This meeting was the litmus test for determining whether AI could complete this project. We printed musical scores that had been developed by AI and built off the sketches from Beethoven’s 10th. A pianist performed in a small concert hall in the museum before a group of journalists, music scholars, and Beethoven experts.

Journalists and musicians gather to hear a pianist perform parts of Beethoven’s 10th Symphony. Image Credit: Ahmed Elgammal, CC BY-SA

We challenged the audience to determine where Beethoven’s phrases ended and where the AI extrapolation began. They couldn’t.

A few days later, one of these AI-generated scores was played by a string quartet in a news conference. Only those who intimately knew Beethoven’s sketches for the 10th Symphony could determine when the AI-generated parts came in.

The success of these tests told us we were on the right track. But these were just a couple of minutes of music. There was still much more work to do.

Ready for the World
At every point, Beethoven’s genius loomed, challenging us to do better. As the project evolved, the AI did as well. Over the ensuing 18 months, we constructed and orchestrated two entire movements of more than 20 minutes apiece.

We anticipate some pushback to this work—those who will say that the arts should be off-limits from AI, and that AI has no business trying to replicate the human creative process. Yet when it comes to the arts, I see AI not as a replacement, but as a tool—one that opens doors for artists to express themselves in new ways.

This project would not have been possible without the expertise of human historians and musicians. It took an immense amount of work—and, yes, creative thinking—to accomplish this goal.

At one point, one of the music experts on the team said that the AI reminded him of an eager music student who practices every day, learns, and becomes better and better.

Now that student, having taken the baton from Beethoven, is ready to present the 10th Symphony to the world.

The piece above is a selection from Beethoven’s 10th Symphony. YouTube/Modern Recordings, CC BY-SA 3.38 MB (download)

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Image Credit: Circe Denyer Continue reading

Posted in Human Robots

#439815 How to Prepare Your Workforce for AI ...

Image by John Conde from Pixabay Despite a myriad of articles, research papers, and conversations regarding artificial intelligence and machine learning development, the predictions about its impact range significantly. The absolute majority agrees that AI is one of the keys to digital transformation and that it will change the business and job market forever. However, it’s …

The post How to Prepare Your Workforce for AI Disruption? appeared first on TFOT. Continue reading

Posted in Human Robots

#439804 How Quantum Computers Can Be Used to ...

Using computer simulations to design new chips played a crucial role in the rapid improvements in processor performance we’ve experienced in recent decades. Now Chinese researchers have extended the approach to the quantum world.

Electronic design automation tools started to become commonplace in the early 1980s as the complexity of processors rose exponentially, and today they are an indispensable tool for chip designers.

More recently, Google has been turbocharging the approach by using artificial intelligence to design the next generation of its AI chips. This holds the promise of setting off a process of recursive self-improvement that could lead to rapid performance gains for AI.

Now, New Scientist has reported on a team from the University of Science and Technology of China in Shanghai that has applied the same ideas to another emerging field of computing: quantum processors. In a paper posted to the arXiv pre-print server, the researchers describe how they used a quantum computer to design a new type of qubit that significantly outperformed their previous design.

“Simulations of high-complexity quantum systems, which are intractable for classical computers, can be efficiently done with quantum computers,” the authors wrote. “Our work opens the way to designing advanced quantum processors using existing quantum computing resources.”

At the heart of the idea is the fact that the complexity of quantum systems grows exponentially as they increase in size. As a result, even the most powerful supercomputers struggle to simulate fairly small quantum systems.

This was the basis for Google’s groundbreaking display of “quantum supremacy” in 2019. The company’s researchers used a 53-qubit processor to run a random quantum circuit a million times and showed that it would take roughly 10,000 years to simulate the experiment on the world’s fastest supercomputer.

This means that using classical computers to help in the design of new quantum computers is likely to hit fundamental limits pretty quickly. Using a quantum computer, however, sidesteps the problem because it can exploit the same oddities of the quantum world that make the problem complex in the first place.

This is exactly what the Chinese researchers did. They used an algorithm called a variational quantum eigensolver to simulate the kind of superconducting electronic circuit found at the heart of a quantum computer. This was used to explore what happens when certain energy levels in the circuit are altered.

Normally this kind of experiment would require them to build large numbers of physical prototypes and test them, but instead the team was able to rapidly model the impact of the changes. The upshot was that the researchers discovered a new type of qubit that was more powerful than the one they were already using.

Any two-level quantum system can act as a qubit, but most superconducting quantum computers use transmons, which encode quantum states into the oscillations of electrons. By tweaking the energy levels of their simulated quantum circuit, the researchers were able to discover a new qubit design they dubbed a plasonium.

It is less than half the size of a transmon, and when the researchers fabricated it they found that it holds its quantum state for longer and is less prone to errors. It still works on similar principles to the transmon, so it’s possible to manipulate it using the same control technologies.

The researchers point out that this is only a first prototype, so with further optimization and the integration of recent progress in new superconducting materials and surface treatment methods they expect performance to increase even more.

But the new qubit the researchers have designed is probably not their most significant contribution. By demonstrating that even today’s rudimentary quantum computers can help design future devices, they’ve opened the door to a virtuous cycle that could significantly speed innovation in this field.

Image Credit: Pete Linforth from Pixabay Continue reading

Posted in Human Robots

#439773 How the U.S. Army Is Turning Robots Into ...

This article is part of our special report on AI, “The Great AI Reckoning.”

“I should probably not be standing this close,” I think to myself, as the robot slowly approaches a large tree branch on the floor in front of me. It's not the size of the branch that makes me nervous—it's that the robot is operating autonomously, and that while I know what it's supposed to do, I'm not entirely sure what it will do. If everything works the way the roboticists at the U.S. Army Research Laboratory (ARL) in Adelphi, Md., expect, the robot will identify the branch, grasp it, and drag it out of the way. These folks know what they're doing, but I've spent enough time around robots that I take a small step backwards anyway.

The robot, named
RoMan, for Robotic Manipulator, is about the size of a large lawn mower, with a tracked base that helps it handle most kinds of terrain. At the front, it has a squat torso equipped with cameras and depth sensors, as well as a pair of arms that were harvested from a prototype disaster-response robot originally developed at NASA's Jet Propulsion Laboratory for a DARPA robotics competition. RoMan's job today is roadway clearing, a multistep task that ARL wants the robot to complete as autonomously as possible. Instead of instructing the robot to grasp specific objects in specific ways and move them to specific places, the operators tell RoMan to “go clear a path.” It's then up to the robot to make all the decisions necessary to achieve that objective.

The ability to make decisions autonomously is not just what makes robots useful, it's what makes robots
robots. We value robots for their ability to sense what's going on around them, make decisions based on that information, and then take useful actions without our input. In the past, robotic decision making followed highly structured rules—if you sense this, then do that. In structured environments like factories, this works well enough. But in chaotic, unfamiliar, or poorly defined settings, reliance on rules makes robots notoriously bad at dealing with anything that could not be precisely predicted and planned for in advance.

RoMan, along with many other robots including home vacuums, drones, and autonomous cars, handles the challenges of semistructured environments through artificial neural networks—a computing approach that loosely mimics the structure of neurons in biological brains. About a decade ago, artificial neural networks began to be applied to a wide variety of semistructured data that had previously been very difficult for computers running rules-based programming (generally referred to as symbolic reasoning) to interpret. Rather than recognizing specific data structures, an artificial neural network is able to recognize data patterns, identifying novel data that are similar (but not identical) to data that the network has encountered before. Indeed, part of the appeal of artificial neural networks is that they are trained by example, by letting the network ingest annotated data and learn its own system of pattern recognition. For neural networks with multiple layers of abstraction, this technique is called deep learning.

Even though humans are typically involved in the training process, and even though artificial neural networks were inspired by the neural networks in human brains, the kind of pattern recognition a deep learning system does is fundamentally different from the way humans see the world. It's often nearly impossible to understand the relationship between the data input into the system and the interpretation of the data that the system outputs. And that difference—the “black box” opacity of deep learning—poses a potential problem for robots like RoMan and for the Army Research Lab.

In chaotic, unfamiliar, or poorly defined settings, reliance on rules makes robots notoriously bad at dealing with anything that could not be precisely predicted and planned for in advance.

This opacity means that robots that rely on deep learning have to be used carefully. A deep-learning system is good at recognizing patterns, but lacks the world understanding that a human typically uses to make decisions, which is why such systems do best when their applications are well defined and narrow in scope. “When you have well-structured inputs and outputs, and you can encapsulate your problem in that kind of relationship, I think deep learning does very well,” says
Tom Howard, who directs the University of Rochester's Robotics and Artificial Intelligence Laboratory and has developed natural-language interaction algorithms for RoMan and other ground robots. “The question when programming an intelligent robot is, at what practical size do those deep-learning building blocks exist?” Howard explains that when you apply deep learning to higher-level problems, the number of possible inputs becomes very large, and solving problems at that scale can be challenging. And the potential consequences of unexpected or unexplainable behavior are much more significant when that behavior is manifested through a 170-kilogram two-armed military robot.

After a couple of minutes, RoMan hasn't moved—it's still sitting there, pondering the tree branch, arms poised like a praying mantis. For the last 10 years, the Army Research Lab's Robotics Collaborative Technology Alliance (RCTA) has been working with roboticists from Carnegie Mellon University, Florida State University, General Dynamics Land Systems, JPL, MIT, QinetiQ North America, University of Central Florida, the University of Pennsylvania, and other top research institutions to develop robot autonomy for use in future ground-combat vehicles. RoMan is one part of that process.

The “go clear a path” task that RoMan is slowly thinking through is difficult for a robot because the task is so abstract. RoMan needs to identify objects that might be blocking the path, reason about the physical properties of those objects, figure out how to grasp them and what kind of manipulation technique might be best to apply (like pushing, pulling, or lifting), and then make it happen. That's a lot of steps and a lot of unknowns for a robot with a limited understanding of the world.

This limited understanding is where the ARL robots begin to differ from other robots that rely on deep learning, says Ethan Stump, chief scientist of the AI for Maneuver and Mobility program at ARL. “The Army can be called upon to operate basically anywhere in the world. We do not have a mechanism for collecting data in all the different domains in which we might be operating. We may be deployed to some unknown forest on the other side of the world, but we'll be expected to perform just as well as we would in our own backyard,” he says. Most deep-learning systems function reliably only within the domains and environments in which they've been trained. Even if the domain is something like “every drivable road in San Francisco,” the robot will do fine, because that's a data set that has already been collected. But, Stump says, that's not an option for the military. If an Army deep-learning system doesn't perform well, they can't simply solve the problem by collecting more data.

ARL's robots also need to have a broad awareness of what they're doing. “In a standard operations order for a mission, you have goals, constraints, a paragraph on the commander's intent—basically a narrative of the purpose of the mission—which provides contextual info that humans can interpret and gives them the structure for when they need to make decisions and when they need to improvise,” Stump explains. In other words, RoMan may need to clear a path quickly, or it may need to clear a path quietly, depending on the mission's broader objectives. That's a big ask for even the most advanced robot. “I can't think of a deep-learning approach that can deal with this kind of information,” Stump says.

Robots at the Army Research Lab test autonomous navigation techniques in rough terrain [top, middle] with the goal of being able to keep up with their human teammates. ARL is also developing robots with manipulation capabilities [bottom] that can interact with objects so that humans don't have to.Evan Ackerman

While I watch, RoMan is reset for a second try at branch removal. ARL's approach to autonomy is modular, where deep learning is combined with other techniques, and the robot is helping ARL figure out which tasks are appropriate for which techniques. At the moment, RoMan is testing two different ways of identifying objects from 3D sensor data: UPenn's approach is deep-learning-based, while Carnegie Mellon is using a method called perception through search, which relies on a more traditional database of 3D models. Perception through search works only if you know exactly which objects you're looking for in advance, but training is much faster since you need only a single model per object. It can also be more accurate when perception of the object is difficult—if the object is partially hidden or upside-down, for example. ARL is testing these strategies to determine which is the most versatile and effective, letting them run simultaneously and compete against each other.

Perception is one of the things that deep learning tends to excel at. “The computer vision community has made crazy progress using deep learning for this stuff,” says Maggie Wigness, a computer scientist at ARL. “We've had good success with some of these models that were trained in one environment generalizing to a new environment, and we intend to keep using deep learning for these sorts of tasks, because it's the state of the art.”

ARL's modular approach might combine several techniques in ways that leverage their particular strengths. For example, a perception system that uses deep-learning-based vision to classify terrain could work alongside an autonomous driving system based on an approach called inverse reinforcement learning, where the model can rapidly be created or refined by observations from human soldiers. Traditional reinforcement learning optimizes a solution based on established reward functions, and is often applied when you're not necessarily sure what optimal behavior looks like. This is less of a concern for the Army, which can generally assume that well-trained humans will be nearby to show a robot the right way to do things. “When we deploy these robots, things can change very quickly,” Wigness says. “So we wanted a technique where we could have a soldier intervene, and with just a few examples from a user in the field, we can update the system if we need a new behavior.” A deep-learning technique would require “a lot more data and time,” she says.

It's not just data-sparse problems and fast adaptation that deep learning struggles with. There are also questions of robustness, explainability, and safety. “These questions aren't unique to the military,” says Stump, “but it's especially important when we're talking about systems that may incorporate lethality.” To be clear, ARL is not currently working on lethal autonomous weapons systems, but the lab is helping to lay the groundwork for autonomous systems in the U.S. military more broadly, which means considering ways in which such systems may be used in the future.

The requirements of a deep network are to a large extent misaligned with the requirements of an Army mission, and that's a problem.

Safety is an obvious priority, and yet there isn't a clear way of making a deep-learning system verifiably safe, according to Stump. “Doing deep learning with safety constraints is a major research effort. It's hard to add those constraints into the system, because you don't know where the constraints already in the system came from. So when the mission changes, or the context changes, it's hard to deal with that. It's not even a data question; it's an architecture question.” ARL's modular architecture, whether it's a perception module that uses deep learning or an autonomous driving module that uses inverse reinforcement learning or something else, can form parts of a broader autonomous system that incorporates the kinds of safety and adaptability that the military requires. Other modules in the system can operate at a higher level, using different techniques that are more verifiable or explainable and that can step in to protect the overall system from adverse unpredictable behaviors. “If other information comes in and changes what we need to do, there's a hierarchy there,” Stump says. “It all happens in a rational way.”

Nicholas Roy, who leads the Robust Robotics Group at MIT and describes himself as “somewhat of a rabble-rouser” due to his skepticism of some of the claims made about the power of deep learning, agrees with the ARL roboticists that deep-learning approaches often can't handle the kinds of challenges that the Army has to be prepared for. “The Army is always entering new environments, and the adversary is always going to be trying to change the environment so that the training process the robots went through simply won't match what they're seeing,” Roy says. “So the requirements of a deep network are to a large extent misaligned with the requirements of an Army mission, and that's a problem.”

Roy, who has worked on abstract reasoning for ground robots as part of the RCTA, emphasizes that deep learning is a useful technology when applied to problems with clear functional relationships, but when you start looking at abstract concepts, it's not clear whether deep learning is a viable approach. “I'm very interested in finding how neural networks and deep learning could be assembled in a way that supports higher-level reasoning,” Roy says. “I think it comes down to the notion of combining multiple low-level neural networks to express higher level concepts, and I do not believe that we understand how to do that yet.” Roy gives the example of using two separate neural networks, one to detect objects that are cars and the other to detect objects that are red. It's harder to combine those two networks into one larger network that detects red cars than it would be if you were using a symbolic reasoning system based on structured rules with logical relationships. “Lots of people are working on this, but I haven't seen a real success that drives abstract reasoning of this kind.”

For the foreseeable future, ARL is making sure that its autonomous systems are safe and robust by keeping humans around for both higher-level reasoning and occasional low-level advice. Humans might not be directly in the loop at all times, but the idea is that humans and robots are more effective when working together as a team. When the most recent phase of the Robotics Collaborative Technology Alliance program began in 2009, Stump says, “we'd already had many years of being in Iraq and Afghanistan, where robots were often used as tools. We've been trying to figure out what we can do to transition robots from tools to acting more as teammates within the squad.”

RoMan gets a little bit of help when a human supervisor points out a region of the branch where grasping might be most effective. The robot doesn't have any fundamental knowledge about what a tree branch actually is, and this lack of world knowledge (what we think of as common sense) is a fundamental problem with autonomous systems of all kinds. Having a human leverage our vast experience into a small amount of guidance can make RoMan's job much easier. And indeed, this time RoMan manages to successfully grasp the branch and noisily haul it across the room.

Turning a robot into a good teammate can be difficult, because it can be tricky to find the right amount of autonomy. Too little and it would take most or all of the focus of one human to manage one robot, which may be appropriate in special situations like explosive-ordnance disposal but is otherwise not efficient. Too much autonomy and you'd start to have issues with trust, safety, and explainability.

“I think the level that we're looking for here is for robots to operate on the level of working dogs,” explains Stump. “They understand exactly what we need them to do in limited circumstances, they have a small amount of flexibility and creativity if they are faced with novel circumstances, but we don't expect them to do creative problem-solving. And if they need help, they fall back on us.”

RoMan is not likely to find itself out in the field on a mission anytime soon, even as part of a team with humans. It's very much a research platform. But the software being developed for RoMan and other robots at ARL, called Adaptive Planner Parameter Learning (APPL), will likely be used first in autonomous driving, and later in more complex robotic systems that could include mobile manipulators like RoMan. APPL combines different machine-learning techniques (including inverse reinforcement learning and deep learning) arranged hierarchically underneath classical autonomous navigation systems. That allows high-level goals and constraints to be applied on top of lower-level programming. Humans can use teleoperated demonstrations, corrective interventions, and evaluative feedback to help robots adjust to new environments, while the robots can use unsupervised reinforcement learning to adjust their behavior parameters on the fly. The result is an autonomy system that can enjoy many of the benefits of machine learning, while also providing the kind of safety and explainability that the Army needs. With APPL, a learning-based system like RoMan can operate in predictable ways even under uncertainty, falling back on human tuning or human demonstration if it ends up in an environment that's too different from what it trained on.

It's tempting to look at the rapid progress of commercial and industrial autonomous systems (autonomous cars being just one example) and wonder why the Army seems to be somewhat behind the state of the art. But as Stump finds himself having to explain to Army generals, when it comes to autonomous systems, “there are lots of hard problems, but industry's hard problems are different from the Army's hard problems.” The Army doesn't have the luxury of operating its robots in structured environments with lots of data, which is why ARL has put so much effort into APPL, and into maintaining a place for humans. Going forward, humans are likely to remain a key part of the autonomous framework that ARL is developing. “That's what we're trying to build with our robotics systems,” Stump says. “That's our bumper sticker: 'From tools to teammates.' ”

This article appears in the October 2021 print issue as “Deep Learning Goes to Boot Camp.”

Special Report: The Great AI Reckoning

READ NEXT:
7 Revealing Ways AIs Fail

Or see the full report for more articles on the future of AI. Continue reading

Posted in Human Robots