Tag Archives: inspiration
Dr. Been Kim wants to rip open the black box of deep learning.
A senior researcher at Google Brain, Kim specializes in a sort of AI psychology. Like cognitive psychologists before her, she develops various ways to probe the alien minds of artificial neural networks (ANNs), digging into their gory details to better understand the models and their responses to inputs.
The more interpretable ANNs are, the reasoning goes, the easier it is to reveal potential flaws in their reasoning. And if we understand when or why our systems choke, we’ll know when not to use them—a foundation for building responsible AI.
There are already several ways to tap into ANN reasoning, but Kim’s inspiration for unraveling the AI black box came from an entirely different field: cognitive psychology. The field aims to discover fundamental rules of how the human mind—essentially also a tantalizing black box—operates, Kim wrote with her colleagues.
In a new paper uploaded to the pre-publication server arXiv, the team described a way to essentially perform a human cognitive test on ANNs. The test probes how we automatically complete gaps in what we see, so that they form entire objects—for example, perceiving a circle from a bunch of loose dots arranged along a clock face. Psychologist dub this the “law of completion,” a highly influential idea that led to explanations of how our minds generalize data into concepts.
Because deep neural networks in machine vision loosely mimic the structure and connections of the visual cortex, the authors naturally asked: do ANNs also exhibit the law of completion? And what does that tell us about how an AI thinks?
Enter the Germans
The law of completion is part of a series of ideas from Gestalt psychology. Back in the 1920s, long before the advent of modern neuroscience, a group of German experimental psychologists asked: in this chaotic, flashy, unpredictable world, how do we piece together input in a way that leads to meaningful perceptions?
The result is a group of principles known together as the Gestalt effect: that the mind self-organizes to form a global whole. In the more famous words of Gestalt psychologist Kurt Koffka, our perception forms a whole that’s “something else than the sum of its parts.” Not greater than; just different.
Although the theory has its critics, subsequent studies in humans and animals suggest that the law of completion happens on both the cognitive and neuroanatomical level.
Take a look at the drawing below. You immediately “see” a shape that’s actually the negative: a triangle or a square (A and B). Or you further perceive a 3D ball (C), or a snake-like squiggle (D). Your mind fills in blank spots, so that the final perception is more than just the black shapes you’re explicitly given.
Image Credit: Wikimedia Commons contributors, the free media repository.
Neuroscientists now think that the effect comes from how our visual system processes information. Arranged in multiple layers and columns, lower-level neurons—those first to wrangle the data—tend to extract simpler features such as lines or angles. In Gestalt speak, they “see” the parts.
Then, layer by layer, perception becomes more abstract, until higher levels of the visual system directly interpret faces or objects—or things that don’t really exist. That is, the “whole” emerges.
The Experiment Setup
Inspired by these classical experiments, Kim and team developed a protocol to test the Gestalt effect on feed-forward ANNs: one simple, the other, dubbed the “Inception V3,” far more complex and widely used in the machine vision community.
The main idea is similar to the triangle drawings above. First, the team generated three datasets: one set shows complete, ordinary triangles. The second—the “Illusory” set, shows triangles with the edges removed but the corners intact. Thanks to the Gestalt effect, to us humans these generally still look like triangles. The third set also only shows incomplete triangle corners. But here, the corners are randomly rotated so that we can no longer imagine a line connecting them—hence, no more triangle.
To generate a dataset large enough to tease out small effects, the authors changed the background color, image rotation, and other aspects of the dataset. In all, they produced nearly 1,000 images to test their ANNs on.
“At a high level, we compare an ANN’s activation similarities between the three sets of stimuli,” the authors explained. The process is two steps: first, train the AI on complete triangles. Second, test them on the datasets. If the response is more similar between the illusory set and the complete triangle—rather than the randomly rotated set—it should suggest a sort of Gestalt closure effect in the network.
Right off the bat, the team got their answer: yes, ANNs do seem to exhibit the law of closure.
When trained on natural images, the networks better classified the illusory set as triangles than those with randomized connection weights or networks trained on white noise.
When the team dug into the “why,” things got more interesting. The ability to complete an image correlated with the network’s ability to generalize.
Humans subconsciously do this constantly: anything with a handle made out of ceramic, regardless of shape, could easily be a mug. ANNs still struggle to grasp common features—clues that immediately tells us “hey, that’s a mug!” But when they do, it sometimes allows the networks to better generalize.
“What we observe here is that a network that is able to generalize exhibits…more of the closure effect [emphasis theirs], hinting that the closure effect reflects something beyond simply learning features,” the team wrote.
What’s more, remarkably similar to the visual cortex, “higher” levels of the ANNs showed more of the closure effect than lower layers, and—perhaps unsurprisingly—the more layers a network had, the more it exhibited the closure effect.
As the networks learned, their ability to map out objects from fragments also improved. When the team messed around with the brightness and contrast of the images, the AI still learned to see the forest from the trees.
“Our findings suggest that neural networks trained with natural images do exhibit closure,” the team concluded.
That’s not to say that ANNs recapitulate the human brain. As Google’s Deep Dream, an effort to coax AIs into spilling what they’re perceiving, clearly demonstrates, machine vision sees some truly weird stuff.
In contrast, because they’re modeled after the human visual cortex, perhaps it’s not all that surprising that these networks also exhibit higher-level properties inherent to how we process information.
But to Kim and her colleagues, that’s exactly the point.
“The field of psychology has developed useful tools and insights to study human brains– tools that we may be able to borrow to analyze artificial neural networks,” they wrote.
By tweaking these tools to better analyze machine minds, the authors were able to gain insight on how similarly or differently they see the world from us. And that’s the crux: the point isn’t to say that ANNs perceive the world sort of, kind of, maybe similar to humans. It’s to tap into a wealth of cognitive psychology tools, established over decades using human minds, to probe that of ANNs.
“The work here is just one step along a much longer path,” the authors conclude.
“Understanding where humans and neural networks differ will be helpful for research on interpretability by enlightening the fundamental differences between the two interesting species.”
Image Credit: Popova Alena / Shutterstock.com Continue reading
According to some scientists, humans really do have a sixth sense. There’s nothing supernatural about it: the sense of proprioception tells you about the relative positions of your limbs and the rest of your body. Close your eyes, block out all sound, and you can still use this internal “map” of your external body to locate your muscles and body parts – you have an innate sense of the distances between them, and the perception of how they’re moving, above and beyond your sense of touch.
This sense is invaluable for allowing us to coordinate our movements. In humans, the brain integrates senses including touch, heat, and the tension in muscle spindles to allow us to build up this map.
Replicating this complex sense has posed a great challenge for roboticists. We can imagine simulating the sense of sight with cameras, sound with microphones, or touch with pressure-pads. Robots with chemical sensors could be far more accurate than us in smell and taste, but building in proprioception, the robot’s sense of itself and its body, is far more difficult, and is a large part of why humanoid robots are so tricky to get right.
Simultaneous localization and mapping (SLAM) software allows robots to use their own senses to build up a picture of their surroundings and environment, but they’d need a keen sense of the position of their own bodies to interact with it. If something unexpected happens, or in dark environments where primary senses are not available, robots can struggle to keep track of their own position and orientation. For human-robot interaction, wearable robotics, and delicate applications like surgery, tiny differences can be extremely important.
In the case of hard robotics, this is generally solved by using a series of strain and pressure sensors in each joint, which allow the robot to determine how its limbs are positioned. That works fine for rigid robots with a limited number of joints, but for softer, more flexible robots, this information is limited. Roboticists are faced with a dilemma: a vast, complex array of sensors for every degree of freedom in the robot’s movement, or limited skill in proprioception?
New techniques, often involving new arrays of sensory material and machine-learning algorithms to fill in the gaps, are starting to tackle this problem. Take the work of Thomas George Thuruthel and colleagues in Pisa and San Diego, who draw inspiration from the proprioception of humans. In a new paper in Science Robotics, they describe the use of soft sensors distributed through a robotic finger at random. This placement is much like the constant adaptation of sensors in humans and animals, rather than relying on feedback from a limited number of positions.
The sensors allow the soft robot to react to touch and pressure in many different locations, forming a map of itself as it contorts into complicated positions. The machine-learning algorithm serves to interpret the signals from the randomly-distributed sensors: as the finger moves around, it’s observed by a motion capture system. After training the robot’s neural network, it can associate the feedback from the sensors with the position of the finger detected in the motion-capture system, which can then be discarded. The robot observes its own motions to understand the shapes that its soft body can take, and translate them into the language of these soft sensors.
“The advantages of our approach are the ability to predict complex motions and forces that the soft robot experiences (which is difficult with traditional methods) and the fact that it can be applied to multiple types of actuators and sensors,” said Michael Tolley of the University of California San Diego. “Our method also includes redundant sensors, which improves the overall robustness of our predictions.”
The use of machine learning lets the roboticists come up with a reliable model for this complex, non-linear system of motions for the actuators, something difficult to do by directly calculating the expected motion of the soft-bot. It also resembles the human system of proprioception, built on redundant sensors that change and shift in position as we age.
In Search of a Perfect Arm
Another approach to training robots in using their bodies comes from Robert Kwiatkowski and Hod Lipson of Columbia University in New York. In their paper “Task-agnostic self-modeling machines,” also recently published in Science Robotics, they describe a new type of robotic arm.
Robotic arms and hands are getting increasingly dexterous, but training them to grasp a large array of objects and perform many different tasks can be an arduous process. It’s also an extremely valuable skill to get right: Amazon is highly interested in the perfect robot arm. Google hooked together an array of over a dozen robot arms so that they could share information about grasping new objects, in part to cut down on training time.
Individually training a robot arm to perform every individual task takes time and reduces the adaptability of your robot: either you need an ML algorithm with a huge dataset of experiences, or, even worse, you need to hard-code thousands of different motions. Kwiatkowski and Lipson attempt to overcome this by developing a robotic system that has a “strong sense of self”: a model of its own size, shape, and motions.
They do this using deep machine learning. The robot begins with no prior knowledge of its own shape or the underlying physics of its motion. It then repeats a series of a thousand random trajectories, recording the motion of its arm. Kwiatkowski and Lipson compare this to a baby in the first year of life observing the motions of its own hands and limbs, fascinated by picking up and manipulating objects.
Again, once the robot has trained itself to interpret these signals and build up a robust model of its own body, it’s ready for the next stage. Using that deep-learning algorithm, the researchers then ask the robot to design strategies to accomplish simple pick-up and place and handwriting tasks. Rather than laboriously and narrowly training itself for each individual task, limiting its abilities to a very narrow set of circumstances, the robot can now strategize how to use its arm for a much wider range of situations, with no additional task-specific training.
In a further experiment, the researchers replaced part of the arm with a “deformed” component, intended to simulate what might happen if the robot was damaged. The robot can then detect that something’s up and “reconfigure” itself, reconstructing its self-model by going through the training exercises once again; it was then able to perform the same tasks with only a small reduction in accuracy.
Machine learning techniques are opening up the field of robotics in ways we’ve never seen before. Combining them with our understanding of how humans and other animals are able to sense and interact with the world around us is bringing robotics closer and closer to becoming truly flexible and adaptable, and, eventually, omnipresent.
But before they can get out and shape the world, as these studies show, they will need to understand themselves.
Image Credit: jumbojan / Shutterstock.com Continue reading