Tag Archives: Visual
As a human you instinctively know that a leopard is closer to a cat than a motorbike, but the way we train most AI makes them oblivious to these kinds of relations. Building the concept of similarity into our algorithms could make them far more capable, writes the author of a new paper in Science Robotics.
Convolutional neural networks have revolutionized the field of computer vision to the point that machines are now outperforming humans on some of the most challenging visual tasks. But the way we train them to analyze images is very different from the way humans learn, says Atsuto Maki, an associate professor at KTH Royal Institute of Technology.
“Imagine that you are two years old and being quizzed on what you see in a photo of a leopard,” he writes. “You might answer ‘a cat’ and your parents might say, ‘yeah, not quite but similar’.”
In contrast, the way we train neural networks rarely gives that kind of partial credit. They are typically trained to have very high confidence in the correct label and consider all incorrect labels, whether ”cat” or “motorbike,” equally wrong. That’s a mistake, says Maki, because ignoring the fact that something can be “less wrong” means you’re not exploiting all of the information in the training data.
Even when models are trained this way, there will be small differences in the probabilities assigned to incorrect labels that can tell you a lot about how well the model can generalize what it has learned to unseen data.
If you show a model a picture of a leopard and it gives “cat” a probability of five percent and “motorbike” one percent, that suggests it picked up on the fact that a cat is closer to a leopard than a motorbike. In contrast, if the figures are the other way around it means the model hasn’t learned the broad features that make cats and leopards similar, something that could potentially be helpful when analyzing new data.
If we could boost this ability to identify similarities between classes we should be able to create more flexible models better able to generalize, says Maki. And recent research has demonstrated how variations of an approach called regularization might help us achieve that goal.
Neural networks are prone to a problem called “overfitting,” which refers to a tendency to pay too much attention to tiny details and noise specific to their training set. When that happens, models will perform excellently on their training data but poorly when applied to unseen test data without these particular quirks.
Regularization is used to circumvent this problem, typically by reducing the network’s capacity to learn all this unnecessary information and therefore boost its ability to generalize to new data. Techniques are varied, but generally involve modifying the network’s structure or the strength of the weights between artificial neurons.
More recently, though, researchers have suggested new regularization approaches that work by encouraging a broader spread of probabilities across all classes. This essentially helps them capture more of the class similarities, says Maki, and therefore boosts their ability to generalize.
One such approach was devised in 2017 by Google Brain researchers, led by deep learning pioneer Geoffrey Hinton. They introduced a penalty to their training process that directly punished overconfident predictions in the model’s outputs, and a technique called label smoothing that prevents the largest probability becoming much larger than all others. This meant the probabilities were lower for correct labels and higher for incorrect ones, which was found to boost performance of models on varied tasks from image classification to speech recognition.
Another came from Maki himself in 2017 and achieves the same goal, but by suppressing high values in the model’s feature vector—the mathematical construct that describes all of an object’s important characteristics. This has a knock-on effect on the spread of output probabilities and also helped boost performance on various image classification tasks.
While it’s still early days for the approach, the fact that humans are able to exploit these kinds of similarities to learn more efficiently suggests that models that incorporate them hold promise. Maki points out that it could be particularly useful in applications such as robotic grasping, where distinguishing various similar objects is important.
Image Credit: Marianna Kalashnyk / Shutterstock.com Continue reading
A few years back, DeepMind’s Demis Hassabis famously prophesized that AI and neuroscience will positively feed into each other in a “virtuous circle.” If realized, this would fundamentally expand our insight into intelligence, both machine and human.
We’ve already seen some proofs of concept, at least in the brain-to-AI direction. For example, memory replay, a biological mechanism that fortifies our memories during sleep, also boosted AI learning when abstractly appropriated into deep learning models. Reinforcement learning, loosely based on our motivation circuits, is now behind some of AI’s most powerful tools.
Hassabis is about to be proven right again.
Last week, two studies independently tapped into the power of ANNs to solve a 70-year-old neuroscience mystery: how does our visual system perceive reality?
The first, published in Cell, used generative networks to evolve DeepDream-like images that hyper-activate complex visual neurons in monkeys. These machine artworks are pure nightmare fuel to the human eye; but together, they revealed a fundamental “visual hieroglyph” that may form a basic rule for how we piece together visual stimuli to process sight into perception.
In the second study, a team used a deep ANN model—one thought to mimic biological vision—to synthesize new patterns tailored to control certain networks of visual neurons in the monkey brain. When directly shown to monkeys, the team found that the machine-generated artworks could reliably activate predicted populations of neurons. Future improved ANN models could allow even better control, giving neuroscientists a powerful noninvasive tool to study the brain. The work was published in Science.
The individual results, though fascinating, aren’t necessarily the point. Rather, they illustrate how scientists are now striving to complete the virtuous circle: tapping AI to probe natural intelligence. Vision is only the beginning—the tools can potentially be expanded into other sensory domains. And the more we understand about natural brains, the better we can engineer artificial ones.
It’s a “great example of leveraging artificial intelligence to study organic intelligence,” commented Dr. Roman Sandler at Kernel.co on Twitter.
ANNs and biological vision have quite the history.
In the late 1950s, the legendary neuroscientist duo David Hubel and Torsten Wiesel became some of the first to use mathematical equations to understand how neurons in the brain work together.
In a series of experiments—many using cats—the team carefully dissected the structure and function of the visual cortex. Using myriads of images, they revealed that vision is processed in a hierarchy: neurons in “earlier” brain regions, those closer to the eyes, tend to activate when they “see” simple patterns such as lines. As we move deeper into the brain, from the early V1 to a nub located slightly behind our ears, the IT cortex, neurons increasingly respond to more complex or abstract patterns, including faces, animals, and objects. The discovery led some scientists to call certain IT neurons “Jennifer Aniston cells,” which fire in response to pictures of the actress regardless of lighting, angle, or haircut. That is, IT neurons somehow extract visual information into the “gist” of things.
That’s not trivial. The complex neural connections that lead to increasing abstraction of what we see into what we think we see—what we perceive—is a central question in machine vision: how can we teach machines to transform numbers encoding stimuli into dots, lines, and angles that eventually form “perceptions” and “gists”? The answer could transform self-driving cars, facial recognition, and other computer vision applications as they learn to better generalize.
Hubel and Wiesel’s Nobel-prize-winning studies heavily influenced the birth of ANNs and deep learning. Much of earlier ANN “feed-forward” model structures are based on our visual system; even today, the idea of increasing layers of abstraction—for perception or reasoning—guide computer scientists to build AI that can better generalize. The early romance between vision and deep learning is perhaps the bond that kicked off our current AI revolution.
It only seems fair that AI would feed back into vision neuroscience.
Hieroglyphs and Controllers
In the Cell study, a team led by Dr. Margaret Livingstone at Harvard Medical School tapped into generative networks to unravel IT neurons’ complex visual alphabet.
Scientists have long known that neurons in earlier visual regions (V1) tend to fire in response to “grating patches” oriented in certain ways. Using a limited set of these patches like letters, V1 neurons can “express a visual sentence” and represent any image, said Dr. Arash Afraz at the National Institute of Health, who was not involved in the study.
But how IT neurons operate remained a mystery. Here, the team used a combination of genetic algorithms and deep generative networks to “evolve” computer art for every studied neuron. In seven monkeys, the team implanted electrodes into various parts of the visual IT region so that they could monitor the activity of a single neuron.
The team showed each monkey an initial set of 40 images. They then picked the top 10 images that stimulated the highest neural activity, and married them to 30 new images to “evolve” the next generation of images. After 250 generations, the technique, XDREAM, generated a slew of images that mashed up contorted face-like shapes with lines, gratings, and abstract shapes.
This image shows the evolution of an optimum image for stimulating a visual neuron in a monkey. Image Credit: Ponce, Xiao, and Schade et al. – Cell.
“The evolved images look quite counter-intuitive,” explained Afraz. Some clearly show detailed structures that resemble natural images, while others show complex structures that can’t be characterized by our puny human brains.
This figure shows natural images (right) and images evolved by neurons in the inferotemporal cortex of a monkey (left). Image Credit: Ponce, Xiao, and Schade et al. – Cell.
“What started to emerge during each experiment were pictures that were reminiscent of shapes in the world but were not actual objects in the world,” said study author Carlos Ponce. “We were seeing something that was more like the language cells use with each other.”
This image was evolved by a neuron in the inferotemporal cortex of a monkey using AI. Image Credit: Ponce, Xiao, and Schade et al. – Cell.
Although IT neurons don’t seem to use a simple letter alphabet, it does rely on a vast array of characters like hieroglyphs or Chinese characters, “each loaded with more information,” said Afraz.
The adaptive nature of XDREAM turns it into a powerful tool to probe the inner workings of our brains—particularly for revealing discrepancies between biology and models.
The Science study, led by Dr. James DiCarlo at MIT, takes a similar approach. Using ANNs to generate new patterns and images, the team was able to selectively predict and independently control neuron populations in a high-level visual region called V4.
“So far, what has been done with these models is predicting what the neural responses would be to other stimuli that they have not seen before,” said study author Dr. Pouya Bashivan. “The main difference here is that we are going one step further and using the models to drive the neurons into desired states.”
It suggests that our current ANN models for visual computation “implicitly capture a great deal of visual knowledge” which we can’t really describe, but which the brain uses to turn vision information into perception, the authors said. By testing AI-generated images on biological vision, however, the team concluded that today’s ANNs have a degree of understanding and generalization. The results could potentially help engineer even more accurate ANN models of biological vision, which in turn could feed back into machine vision.
“One thing is clear already: Improved ANN models … have led to control of a high-level neural population that was previously out of reach,” the authors said. “The results presented here have likely only scratched the surface of what is possible with such implemented characterizations of the brain’s neural networks.”
To Afraz, the power of AI here is to find cracks in human perception—both our computational models of sensory processes, as well as our evolved biological software itself. AI can be used “as a perfect adversarial tool to discover design cracks” of IT, said Afraz, such as finding computer art that “fools” a neuron into thinking the object is something else.
“As artificial intelligence researchers develop models that work as well as the brain does—or even better—we will still need to understand which networks are more likely to behave safely and further human goals,” said Ponce. “More efficient AI can be grounded by knowledge of how the brain works.”
Image Credit: Sangoiri / Shutterstock.com Continue reading
Robots have been masters of manufacturing at speed and precision for decades, but give them a seemingly simple task like stacking shelves, and they quickly get stuck. That’s changing, though, as engineers build systems that can take on the deceptively tricky tasks most humans can do with their eyes closed.
Boston Dynamics is famous for dramatic reveals of robots performing mind-blowing feats that also leave you scratching your head as to what the market is—think the bipedal Atlas doing backflips or Spot the galloping robot dog.
Last week, the company released a video of a robot called Handle that looks like an ostrich on wheels carrying out the seemingly mundane task of stacking boxes in a warehouse.
It might seem like a step backward, but this is exactly the kind of practical task robots have long struggled with. While the speed and precision of industrial robots has seen them take over many functions in modern factories, they’re generally limited to highly prescribed tasks carried out in meticulously-controlled environments.
That’s because despite their mechanical sophistication, most are still surprisingly dumb. They can carry out precision welding on a car or rapidly assemble electronics, but only by rigidly following a prescribed set of motions. Moving cardboard boxes around a warehouse might seem simple to a human, but it actually involves a variety of tasks machines still find pretty difficult—perceiving your surroundings, navigating, and interacting with objects in a dynamic environment.
But the release of this video suggests Boston Dynamics thinks these kinds of applications are close to prime time. Last week the company doubled down by announcing the acquisition of start-up Kinema Systems, which builds computer vision systems for robots working in warehouses.
It’s not the only company making strides in this area. On the same day the video went live, Google unveiled a robot arm called TossingBot that can pick random objects from a box and quickly toss them into another container beyond its reach, which could prove very useful for sorting items in a warehouse. The machine can train on new objects in just an hour or two, and can pick and toss up to 500 items an hour with better accuracy than any of the humans who tried the task.
And an apple-picking robot built by Abundant Robotics is currently on New Zealand farms navigating between rows of apple trees using LIDAR and computer vision to single out ripe apples before using a vacuum tube to suck them off the tree.
In most cases, advances in machine learning and computer vision brought about by the recent AI boom are the keys to these rapidly improving capabilities. Robots have historically had to be painstakingly programmed by humans to solve each new task, but deep learning is making it possible for them to quickly train themselves on a variety of perception, navigation, and dexterity tasks.
It’s not been simple, though, and the application of deep learning in robotics has lagged behind other areas. A major limitation is that the process typically requires huge amounts of training data. That’s fine when you’re dealing with image classification, but when that data needs to be generated by real-world robots it can make the approach impractical. Simulations offer the possibility to run this training faster than real time, but it’s proved difficult to translate policies learned in virtual environments into the real world.
Recent years have seen significant progress on these fronts, though, and the increasing integration of modern machine learning with robotics. In October, OpenAI imbued a robotic hand with human-level dexterity by training an algorithm in a simulation using reinforcement learning before transferring it to the real-world device. The key to ensuring the translation went smoothly was injecting random noise into the simulation to mimic some of the unpredictability of the real world.
And just a couple of weeks ago, MIT researchers demonstrated a new technique that let a robot arm learn to manipulate new objects with far less training data than is usually required. By getting the algorithm to focus on a few key points on the object necessary for picking it up, the system could learn to pick up a previously unseen object after seeing only a few dozen examples (rather than the hundreds or thousands typically required).
How quickly these innovations will trickle down to practical applications remains to be seen, but a number of startups as well as logistics behemoth Amazon are developing robots designed to flexibly pick and place the wide variety of items found in your average warehouse.
Whether the economics of using robots to replace humans at these kinds of menial tasks makes sense yet is still unclear. The collapse of collaborative robotics pioneer Rethink Robotics last year suggests there are still plenty of challenges.
But at the same time, the number of robotic warehouses is expected to leap from 4,000 today to 50,000 by 2025. It may not be long until robots are muscling in on tasks we’ve long assumed only humans could do.
Image Credit: Visual Generation / Shutterstock.com Continue reading
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