Research Area: Computational Neuroscience

​machine learning, modeling, artificial intelligence, neural networks, memory, ​visual intelligence, audition, natural language processing, learning theory

School of Science announces 2024 Infinite Expansion Awards

Anne Trafton |

The MIT School of Science has announced nine postdocs and research scientists as recipients of the 2024 Infinite Expansion Award, which highlights extraordinary members of the MIT community.

The following are the 2024 School of Science Infinite Expansion winners:

  • Sarthak Chandra, a research scientist in the Department of Brain and Cognitive Sciences, was nominated by Professor Ila Fiete, who wrote, “He has expanded the research abilities of my group by being a versatile and brilliant scientist, by drawing connections with a different area that he was an expert in from his PhD training, and by being a highly involved and caring mentor.”
  • Michal Fux, a research scientist in the Department of Brain and Cognitive Sciences, was nominated by Professor Pawan Sinha, who wrote, “She is one of those figurative beams of light that not only brilliantly illuminate scientific questions, but also enliven a research team.”
  • Andrew Savinov, a postdoc in the Department of Biology, was nominated by Associate Professor Gene-Wei Li, who wrote, “Andrew is an extraordinarily creative and accomplished biophysicist, as well as an outstanding contributor to the broader MIT community.”
  • Ho Fung Cheng, a postdoc in the Department of Chemistry, was nominated by Professor Jeremiah Johnson, who wrote, “His impact on research and our departmental community during his time at MIT has been outstanding, and I believe that he will be a worldclass teacher and research group leader in his independent career next year.”
  • Gabi Wenzel, a postdoc in the Department of Chemistry, was nominated by Assistant Professor Brett McGuire, who wrote, “In the one year since Gabi joined our team, she has become an indispensable leader, demonstrating exceptional skill, innovation, and dedication in our challenging research environment.”
  • Yu-An Zhang, a postdoc in the Department of Chemistry, was nominated by Professor Alison Wendlandt, who wrote, “He is a creative, deep-thinking scientist and a superb organic chemist. But above all, he is an off-scale mentor and a cherished coworker.”
  • Wouter Van de Pontseele, a senior postdoc in the Laboratory for Nuclear Science, was nominated by Professor Joseph Formaggio, who wrote, “He is a talented scientist with an intense creativity, scholarship, and student mentorship record. In the time he has been with my group, he has led multiple facets of my experimental program and has been a wonderful citizen of the MIT community.”
  • Alexander Shvonski, a lecturer in the Department of Physics, was nominated by Assistant Professor Andrew Vanderburg, who wrote, “… I have been blown away by Alex’s knowledge of education research and best practices, his skills as a teacher and course content designer, and I have been extremely grateful for his assistance.”
  • David Stoppel, a research scientist in The Picower Institute for Learning and Memory, was nominated by Professor Mark Bear and his research group, who wrote, “As impressive as his research achievements might be, David’s most genuine qualification for this award is his incredible commitment to mentorship and the dissemination of knowledge.”

Winners are honored with a monetary award and will be celebrated with family, friends, and nominators at a later date, along with recipients of the Infinite Mile Award.

Exposure to different kinds of music influences how the brain interprets rhythm

Anne Trafton |

When listening to music, the human brain appears to be biased toward hearing and producing rhythms composed of simple integer ratios — for example, a series of four beats separated by equal time intervals (forming a 1:1:1 ratio).

However, the favored ratios can vary greatly between different societies, according to a large-scale study led by researchers at MIT and the Max Planck Institute for Empirical Aesthetics and carried out in 15 countries. The study included 39 groups of participants, many of whom came from societies whose traditional music contains distinctive patterns of rhythm not found in Western music.

“Our study provides the clearest evidence yet for some degree of universality in music perception and cognition, in the sense that every single group of participants that was tested exhibits biases for integer ratios. It also provides a glimpse of the variation that can occur across cultures, which can be quite substantial,” says Nori Jacoby, the study’s lead author and a former MIT postdoc, who is now a research group leader at the Max Planck Institute for Empirical Aesthetics in Frankfurt, Germany.

The brain’s bias toward simple integer ratios may have evolved as a natural error-correction system that makes it easier to maintain a consistent body of music, which human societies often use to transmit information.

“When people produce music, they often make small mistakes. Our results are consistent with the idea that our mental representation is somewhat robust to those mistakes, but it is robust in a way that pushes us toward our preexisting ideas of the structures that should be found in music,” says Josh McDermott, an associate professor of brain and cognitive sciences at MIT and a member of MIT’s McGovern Institute for Brain Research and Center for Brains, Minds, and Machines.

McDermott is the senior author of the study, which appears today in Nature Human Behaviour. The research team also included scientists from more than two dozen institutions around the world.

A global approach

The new study grew out of a smaller analysis that Jacoby and McDermott published in 2017. In that paper, the researchers compared rhythm perception in groups of listeners from the United States and the Tsimane’, an Indigenous society located in the Bolivian Amazon rainforest.

pitch perception study
Nori Jacoby, a former MIT postdoc now at the Max Planck Institute for Empirical Aesthetics, runs an experiment with a member of the Tsimane’ tribe, who have had little exposure to Western music. Photo: Josh McDermott

To measure how people perceive rhythm, the researchers devised a task in which they play a randomly generated series of four beats and then ask the listener to tap back what they heard. The rhythm produced by the listener is then played back to the listener, and they tap it back again. Over several iterations, the tapped sequences became dominated by the listener’s internal biases, also known as priors.

“The initial stimulus pattern is random, but at each iteration the pattern is pushed by the listener’s biases, such that it tends to converge to a particular point in the space of possible rhythms,” McDermott says. “That can give you a picture of what we call the prior, which is the set of internal implicit expectations for rhythms that people have in their heads.”

When the researchers first did this experiment, with American college students as the test subjects, they found that people tended to produce time intervals that are related by simple integer ratios. Furthermore, most of the rhythms they produced, such as those with ratios of 1:1:2 and 2:3:3, are commonly found in Western music.

The researchers then went to Bolivia and asked members of the Tsimane’ society to perform the same task. They found that Tsimane’ also produced rhythms with simple integer ratios, but their preferred ratios were different and appeared to be consistent with those that have been documented in the few existing records of Tsimane’ music.

“At that point, it provided some evidence that there might be very widespread tendencies to favor these small integer ratios, and that there might be some degree of cross-cultural variation. But because we had just looked at this one other culture, it really wasn’t clear how this was going to look at a broader scale,” Jacoby says.

To try to get that broader picture, the MIT team began seeking collaborators around the world who could help them gather data on a more diverse set of populations. They ended up studying listeners from 39 groups, representing 15 countries on five continents — North America, South America, Europe, Africa, and Asia.

“This is really the first study of its kind in the sense that we did the same experiment in all these different places, with people who are on the ground in those locations,” McDermott says. “That hasn’t really been done before at anything close to this scale, and it gave us an opportunity to see the degree of variation that might exist around the world.”

A grid of nine different photos showing a researcher working with an individual at a table. The individuals are wearing headphones.
Example testing sites. a, Yaranda, Bolivia. b, Montevideo, Uruguay. c, Sagele, Mali. d, Spitzkoppe, Namibia. e, Pleven, Bulgaria. f, Bamako, Mali. g, D’Kar, Botswana. h, Stockholm, Sweden. i, Guizhou, China. j, Mumbai, India. Verbal informed consent was obtained from the individuals in each photo.

Cultural comparisons

Just as they had in their original 2017 study, the researchers found that in every group they tested, people tended to be biased toward simple integer ratios of rhythm. However, not every group showed the same biases. People from North America and Western Europe, who have likely been exposed to the same kinds of music, were more likely to generate rhythms with the same ratios. However, many groups, for example those in Turkey, Mali, Bulgaria, and Botswana showed a bias for other rhythms.

“There are certain cultures where there are particular rhythms that are prominent in their music, and those end up showing up in the mental representation of rhythm,” Jacoby says.

The researchers believe their findings reveal a mechanism that the brain uses to aid in the perception and production of music.

“When you hear somebody playing something and they have errors in their performance, you’re going to mentally correct for those by mapping them onto where you implicitly think they ought to be,” McDermott says. “If you didn’t have something like this, and you just faithfully represented what you heard, these errors might propagate and make it much harder to maintain a musical system.”

Among the groups that they studied, the researchers took care to include not only college students, who are easy to study in large numbers, but also people living in traditional societies, who are more difficult to reach. Participants from those more traditional groups showed significant differences from college students living in the same countries, and from people who live in those countries but performed the test online.

“What’s very clear from the paper is that if you just look at the results from undergraduate students around the world, you vastly underestimate the diversity that you see otherwise,” Jacoby says. “And the same was true of experiments where we tested groups of people online in Brazil and India, because you’re dealing with people who have internet access and presumably have more exposure to Western music.”

The researchers now hope to run additional studies of different aspects of music perception, taking this global approach.

“If you’re just testing college students around the world or people online, things look a lot more homogenous. I think it’s very important for the field to realize that you actually need to go out into communities and run experiments there, as opposed to taking the low-hanging fruit of running studies with people in a university or on the internet,” McDermott says.

The research was funded by the James S. McDonnell Foundation, the Canadian National Science and Engineering Research Council, the South African National Research Foundation, the United States National Science Foundation, the Chilean National Research and Development Agency, the Austrian Academy of Sciences, the Japan Society for the Promotion of Science, the Keio Global Research Institute, the United Kingdom Arts and Humanities Research Council, the Swedish Research Council, and the John Fell Fund.

The brain runs an internal simulation to keep track of time

by Jennifer Michalowski |

Clocks, computers, and metronomes can keep time with exquisite precision. But even in the absence of an external time keeper, we can track time on our own. We know when minutes or hours have elapsed, and we can maintain a rhythm when we dance, sing, or play music. Now, neuroscientists at the National Autonomous University of Mexico and MIT’s McGovern Institute and have discovered one way the brain keeps a beat: It runs an internal simulation, mentally recreating the perception of an external rhythm and preparing an appropriately timed response.

The discovery, reported January 10, 2024, in the journal Science Advances, illustrates how animals can think about imaginary events and use an internal model to guide their interactions with the world. “It’s a real indication of mental states as an independent driver of behavior,” says neuroscientist Mehrdad Jazayeri, an investigator at the McGovern Institute and an associate professor of brain and cognitive sciences at MIT.

Predicting the future

Jazayeri teamed up with Victor de Lafuente, a neuroscientist at the National Autonomous University of Mexico, to investigate the brain’s time-keeping ability. De Lafuente, who led the study, says they were motivated by curiosity about how the brain makes predictions and prepares for future states of the world.

De Lafuente and his team used a visual metronome to teach monkeys a simple rhythm, showing them a circle that moved between two positions on a screen to set a steady tempo. Then the metronome stopped. After a variable and unpredictable pause, the monkeys were asked to indicate where the dot would be if the metronome had carried on.

Monkeys do well at this task, successfully keeping time after the metronome stops. After the waiting period, they are usually able to identify the expected position of the circle, which they communicate by reaching towards a touchscreen.

To find out how the animals were keeping track of the metronome’s rhythm, de Lafuente’s group monitored their brain activity. In several key brain regions, they found rhythmic patterns of activity that oscillated at the same frequency as the metronome. This occurred while the monkeys watched the metronome. More remarkably, it continued after the metronome had stopped.

“The animal is seeing things going and then things stop. What we find in the brain is the continuation of that process in the animal’s mind,” Jazayeri says. “An entire network is replicating what it was doing.”

That was true in the visual cortex, where clusters of neurons respond to stimuli in specific spots within the eyes’ field of view. One set of cells in the visual cortex fired when the metronome’s circle was on the left of the screen; another set fired when the dot was on the right. As a monkey followed the visual metronome, the researchers could see these cells’ activity alternating rhythmically, tracking the movement. When the metronome stopped, the back-and-forth neural activity continued, maintaining the rhythm. “Once the stimulus was no longer visible, they were seeing the stimulus within their minds,” de Lafuente says.

They found something similar in the brain’s motor cortex, where movements are prepared and executed. De Lafuente explains that the monkeys are motionless for most of their time-keeping task; only when they are asked to indicate where the metronome’s circle should be do they move a hand to touch the screen. But the motor cortex was engaged even before it was time to move. “Within their brains there is a signal that is switching from the left to the right,” he says. “So the monkeys are thinking ‘left, right, left, right’—even when they are not moving and the world is constant.”

While some scientists have proposed that the brain may have a central time-keeping mechanism, the team’s findings indicate that entire networks can be called on to track the passage of time. The monkeys’ model of the future was surprisingly explicit, de Lafuente says, representing specific sensory stimuli and plans for movement. “This offers a potential solution to mentally tracking the dynamics in the world, which is to basically think about them in terms of how they actually would have happened,” Jazayeri says.

 

Complex, unfamiliar sentences make the brain’s language network work harder

Anne Trafton |

With help from an artificial language network, MIT neuroscientists have discovered what kind of sentences are most likely to fire up the brain’s key language processing centers.

The new study reveals that sentences that are more complex, either because of unusual grammar or unexpected meaning, generate stronger responses in these language processing centers. Sentences that are very straightforward barely engage these regions, and nonsensical sequences of words don’t do much for them either.

For example, the researchers found this brain network was most active when reading unusual sentences such as “Buy sell signals remains a particular,” taken from a publicly available language dataset called C4. However, it went quiet when reading something very straightforward, such as “We were sitting on the couch.”

“The input has to be language-like enough to engage the system,” says Evelina Fedorenko, Associate Professor of Neuroscience at MIT and a member of MIT’s McGovern Institute for Brain Research. “And then within that space, if things are really easy to process, then you don’t have much of a response. But if things get difficult, or surprising, if there’s an unusual construction or an unusual set of words that you’re maybe not very familiar with, then the network has to work harder.”

Fedorenko is the senior author of the study, which appears today in Nature Human Behavior. MIT graduate student Greta Tuckute is the lead author of the paper.

Processing language

In this study, the researchers focused on language-processing regions found in the left hemisphere of the brain, which includes Broca’s area as well as other parts of the left frontal and temporal lobes of the brain.

“This language network is highly selective to language, but it’s been harder to actually figure out what is going on in these language regions,” Tuckute says. “We wanted to discover what kinds of sentences, what kinds of linguistic input, drive the left hemisphere language network.”

The researchers began by compiling a set of 1,000 sentences taken from a wide variety of sources — fiction, transcriptions of spoken words, web text, and scientific articles, among many others.

Five human participants read each of the sentences while the researchers measured their language network activity using functional magnetic resonance imaging (fMRI). The researchers then fed those same 1,000 sentences into a large language model — a model similar to ChatGPT, which learns to generate and understand language from predicting the next word in huge amounts of text — and measured the activation patterns of the model in response to each sentence.

Once they had all of those data, the researchers trained a mapping model, known as an “encoding model,” which relates the activation patterns seen in the human brain with those observed in the artificial language model. Once trained, the model could predict how the human language network would respond to any new sentence based on how the artificial language network responded to these 1,000 sentences.

The researchers then used the encoding model to identify 500 new sentences that would generate maximal activity in the human brain (the “drive” sentences), as well as sentences that would elicit minimal activity in the brain’s language network (the “suppress” sentences).

In a group of three new human participants, the researchers found these new sentences did indeed drive and suppress brain activity as predicted.

“This ‘closed-loop’ modulation of brain activity during language processing is novel,” Tuckute says. “Our study shows that the model we’re using (that maps between language-model activations and brain responses) is accurate enough to do this. This is the first demonstration of this approach in brain areas implicated in higher-level cognition, such as the language network.”

Linguistic complexity

To figure out what made certain sentences drive activity more than others, the researchers analyzed the sentences based on 11 different linguistic properties, including grammaticality, plausibility, emotional valence (positive or negative), and how easy it is to visualize the sentence content.

For each of those properties, the researchers asked participants from crowd-sourcing platforms to rate the sentences. They also used a computational technique to quantify each sentence’s “surprisal,” or how uncommon it is compared to other sentences.

This analysis revealed that sentences with higher surprisal generate higher responses in the brain. This is consistent with previous studies showing people have more difficulty processing sentences with higher surprisal, the researchers say.

Another linguistic property that correlated with the language network’s responses was linguistic complexity, which is measured by how much a sentence adheres to the rules of English grammar and how plausible it is, meaning how much sense the content makes, apart from the grammar.

Sentences at either end of the spectrum — either extremely simple, or so complex that they make no sense at all — evoked very little activation in the language network. The largest responses came from sentences that make some sense but require work to figure them out, such as “Jiffy Lube of — of therapies, yes,” which came from the Corpus of Contemporary American English dataset.

“We found that the sentences that elicit the highest brain response have a weird grammatical thing and/or a weird meaning,” Fedorenko says. “There’s something slightly unusual about these sentences.”

The researchers now plan to see if they can extend these findings in speakers of languages other than English. They also hope to explore what type of stimuli may activate language processing regions in the brain’s right hemisphere.

The research was funded by an Amazon Fellowship from the Science Hub, an International Doctoral Fellowship from the American Association of University Women, the MIT-IBM Watson AI Lab, the National Institutes of Health, the McGovern Institute, the Simons Center for the Social Brain, and MIT’s Department of Brain and Cognitive Sciences.

Deep neural networks show promise as models of human hearing

Anne Trafton |

Computational models that mimic the structure and function of the human auditory system could help researchers design better hearing aids, cochlear implants, and brain-machine interfaces. A new study from MIT has found that modern computational models derived from machine learning are moving closer to this goal.

In the largest study yet of deep neural networks that have been trained to perform auditory tasks, the MIT team showed that most of these models generate internal representations that share properties of representations seen in the human brain when people are listening to the same sounds.

The study also offers insight into how to best train this type of model: The researchers found that models trained on auditory input including background noise more closely mimic the activation patterns of the human auditory cortex.

“What sets this study apart is it is the most comprehensive comparison of these kinds of models to the auditory system so far. The study suggests that models that are derived from machine learning are a step in the right direction, and it gives us some clues as to what tends to make them better models of the brain,” says Josh McDermott, an associate professor of brain and cognitive sciences at MIT, a member of MIT’s McGovern Institute for Brain Research and Center for Brains, Minds, and Machines, and the senior author of the study.

MIT graduate student Greta Tuckute and Jenelle Feather PhD ’22 are the lead authors of the open-access paper, which appears today in PLOS Biology.

Models of hearing

Deep neural networks are computational models that consists of many layers of information-processing units that can be trained on huge volumes of data to perform specific tasks. This type of model has become widely used in many applications, and neuroscientists have begun to explore the possibility that these systems can also be used to describe how the human brain performs certain tasks.

“These models that are built with machine learning are able to mediate behaviors on a scale that really wasn’t possible with previous types of models, and that has led to interest in whether or not the representations in the models might capture things that are happening in the brain,” Tuckute says.

When a neural network is performing a task, its processing units generate activation patterns in response to each audio input it receives, such as a word or other type of sound. Those model representations of the input can be compared to the activation patterns seen in fMRI brain scans of people listening to the same input.

In 2018, McDermott and then-graduate student Alexander Kell reported that when they trained a neural network to perform auditory tasks (such as recognizing words from an audio signal), the internal representations generated by the model showed similarity to those seen in fMRI scans of people listening to the same sounds.

Since then, these types of models have become widely used, so McDermott’s research group set out to evaluate a larger set of models, to see if the ability to approximate the neural representations seen in the human brain is a general trait of these models.

For this study, the researchers analyzed nine publicly available deep neural network models that had been trained to perform auditory tasks, and they also created 14 models of their own, based on two different architectures. Most of these models were trained to perform a single task — recognizing words, identifying the speaker, recognizing environmental sounds, and identifying musical genre — while two of them were trained to perform multiple tasks.

When the researchers presented these models with natural sounds that had been used as stimuli in human fMRI experiments, they found that the internal model representations tended to exhibit similarity with those generated by the human brain. The models whose representations were most similar to those seen in the brain were models that had been trained on more than one task and had been trained on auditory input that included background noise.

“If you train models in noise, they give better brain predictions than if you don’t, which is intuitively reasonable because a lot of real-world hearing involves hearing in noise, and that’s plausibly something the auditory system is adapted to,” Feather says.

Hierarchical processing

The new study also supports the idea that the human auditory cortex has some degree of hierarchical organization, in which processing is divided into stages that support distinct computational functions. As in the 2018 study, the researchers found that representations generated in earlier stages of the model most closely resemble those seen in the primary auditory cortex, while representations generated in later model stages more closely resemble those generated in brain regions beyond the primary cortex.

Additionally, the researchers found that models that had been trained on different tasks were better at replicating different aspects of audition. For example, models trained on a speech-related task more closely resembled speech-selective areas.

“Even though the model has seen the exact same training data and the architecture is the same, when you optimize for one particular task, you can see that it selectively explains specific tuning properties in the brain,” Tuckute says.

McDermott’s lab now plans to make use of their findings to try to develop models that are even more successful at reproducing human brain responses. In addition to helping scientists learn more about how the brain may be organized, such models could also be used to help develop better hearing aids, cochlear implants, and brain-machine interfaces.

“A goal of our field is to end up with a computer model that can predict brain responses and behavior. We think that if we are successful in reaching that goal, it will open a lot of doors,” McDermott says.

The research was funded by the National Institutes of Health, an Amazon Fellowship from the Science Hub, an International Doctoral Fellowship from the American Association of University Women, an MIT Friends of McGovern Institute Fellowship, a fellowship from the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center at MIT, and a Department of Energy Computational Science Graduate Fellowship.

What does the future hold for generative AI?

by Adam Zewe |

Speaking at the “Generative AI: Shaping the Future” symposium on Nov. 28, the kickoff event of MIT’s Generative AI Week, keynote speaker and iRobot co-founder Rodney Brooks warned attendees against uncritically overestimating the capabilities of this emerging technology, which underpins increasingly powerful tools like OpenAI’s ChatGPT and Google’s Bard.

“Hype leads to hubris, and hubris leads to conceit, and conceit leads to failure,” cautioned Brooks, who is also a professor emeritus at MIT, a former director of the Computer Science and Artificial Intelligence Laboratory (CSAIL), and founder of Robust.AI.

“No one technology has ever surpassed everything else,” he added.

The symposium, which drew hundreds of attendees from academia and industry to the Institute’s Kresge Auditorium, was laced with messages of hope about the opportunities generative AI offers for making the world a better place, including through art and creativity, interspersed with cautionary tales about what could go wrong if these AI tools are not developed responsibly.

Generative AI is a term to describe machine-learning models that learn to generate new material that looks like the data they were trained on. These models have exhibited some incredible capabilities, such as the ability to produce human-like creative writing, translate languages, generate functional computer code, or craft realistic images from text prompts.

In her opening remarks to launch the symposium, MIT President Sally Kornbluth highlighted several projects faculty and students have undertaken to use generative AI to make a positive impact in the world. For example, the work of the Axim Collaborative, an online education initiative launched by MIT and Harvard, includes exploring the educational aspects of generative AI to help underserved students.

The Institute also recently announced seed grants for 27 interdisciplinary faculty research projects centered on how AI will transform people’s lives across society.

In hosting Generative AI Week, MIT hopes to not only showcase this type of innovation, but also generate “collaborative collisions” among attendees, Kornbluth said.

Collaboration involving academics, policymakers, and industry will be critical if we are to safely integrate a rapidly evolving technology like generative AI in ways that are humane and help humans solve problems, she told the audience.

“I honestly cannot think of a challenge more closely aligned with MIT’s mission. It is a profound responsibility, but I have every confidence that we can face it, if we face it head on and if we face it as a community,” she said.

While generative AI holds the potential to help solve some of the planet’s most pressing problems, the emergence of these powerful machine learning models has blurred the distinction between science fiction and reality, said CSAIL Director Daniela Rus in her opening remarks. It is no longer a question of whether we can make machines that produce new content, she said, but how we can use these tools to enhance businesses and ensure sustainability. 

“Today, we will discuss the possibility of a future where generative AI does not just exist as a technological marvel, but stands as a source of hope and a force for good,” said Rus, who is also the Andrew and Erna Viterbi Professor in the Department of Electrical Engineering and Computer Science.

But before the discussion dove deeply into the capabilities of generative AI, attendees were first asked to ponder their humanity, as MIT Professor Joshua Bennett read an original poem.

Bennett, a professor in the MIT Literature Section and Distinguished Chair of the Humanities, was asked to write a poem about what it means to be human, and drew inspiration from his daughter, who was born three weeks ago.

The poem told of his experiences as a boy watching Star Trek with his father and touched on the importance of passing traditions down to the next generation.

In his keynote remarks, Brooks set out to unpack some of the deep, scientific questions surrounding generative AI, as well as explore what the technology can tell us about ourselves.

To begin, he sought to dispel some of the mystery swirling around generative AI tools like ChatGPT by explaining the basics of how this large language model works. ChatGPT, for instance, generates text one word at a time by determining what the next word should be in the context of what it has already written. While a human might write a story by thinking about entire phrases, ChatGPT only focuses on the next word, Brooks explained.

ChatGPT 3.5 is built on a machine-learning model that has 175 billion parameters and has been exposed to billions of pages of text on the web during training. (The newest iteration, ChatGPT 4, is even larger.) It learns correlations between words in this massive corpus of text and uses this knowledge to propose what word might come next when given a prompt.

The model has demonstrated some incredible capabilities, such as the ability to write a sonnet about robots in the style of Shakespeare’s famous Sonnet 18. During his talk, Brooks showcased the sonnet he asked ChatGPT to write side-by-side with his own sonnet.

But while researchers still don’t fully understand exactly how these models work, Brooks assured the audience that generative AI’s seemingly incredible capabilities are not magic, and it doesn’t mean these models can do anything.

His biggest fears about generative AI don’t revolve around models that could someday surpass human intelligence. Rather, he is most worried about researchers who may throw away decades of excellent work that was nearing a breakthrough, just to jump on shiny new advancements in generative AI; venture capital firms that blindly swarm toward technologies that can yield the highest margins; or the possibility that a whole generation of engineers will forget about other forms of software and AI.

At the end of the day, those who believe generative AI can solve the world’s problems and those who believe it will only generate new problems have at least one thing in common: Both groups tend to overestimate the technology, he said.

“What is the conceit with generative AI? The conceit is that it is somehow going to lead to artificial general intelligence. By itself, it is not,” Brooks said.

Following Brooks’ presentation, a group of MIT faculty spoke about their work using generative AI and participated in a panel discussion about future advances, important but underexplored research topics, and the challenges of AI regulation and policy.

The panel consisted of Jacob Andreas, an associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS) and a member of CSAIL; Antonio Torralba, the Delta Electronics Professor of EECS and a member of CSAIL; Ev Fedorenko, an associate professor of brain and cognitive sciences and an investigator at the McGovern Institute for Brain Research at MIT; and Armando Solar-Lezama, a Distinguished Professor of Computing and associate director of CSAIL. It was moderated by William T. Freeman, the Thomas and Gerd Perkins Professor of EECS and a member of CSAIL.

The panelists discussed several potential future research directions around generative AI, including the possibility of integrating perceptual systems, drawing on human senses like touch and smell, rather than focusing primarily on language and images. The researchers also spoke about the importance of engaging with policymakers and the public to ensure generative AI tools are produced and deployed responsibly.

“One of the big risks with generative AI today is the risk of digital snake oil. There is a big risk of a lot of products going out that claim to do miraculous things but in the long run could be very harmful,” Solar-Lezama said.

The morning session concluded with an excerpt from the 1925 science fiction novel “Metropolis,” read by senior Joy Ma, a physics and theater arts major, followed by a roundtable discussion on the future of generative AI. The discussion included Joshua Tenenbaum, a professor in the Department of Brain and Cognitive Sciences and a member of CSAIL; Dina Katabi, the Thuan and Nicole Pham Professor in EECS and a principal investigator in CSAIL and the MIT Jameel Clinic; and Max Tegmark, professor of physics; and was moderated by Daniela Rus.

One focus of the discussion was the possibility of developing generative AI models that can go beyond what we can do as humans, such as tools that can sense someone’s emotions by using electromagnetic signals to understand how a person’s breathing and heart rate are changing.

But one key to integrating AI like this into the real world safely is to ensure that we can trust it, Tegmark said. If we know an AI tool will meet the specifications we insist on, then “we no longer have to be afraid of building really powerful systems that go out and do things for us in the world,” he said.

The brain may learn about the world the same way some computational models do

Anne Trafton |

To make our way through the world, our brain must develop an intuitive understanding of the physical world around us, which we then use to interpret sensory information coming into the brain.

How does the brain develop that intuitive understanding? Many scientists believe that it may use a process similar to what’s known as “self-supervised learning.” This type of machine learning, originally developed as a way to create more efficient models for computer vision, allows computational models to learn about visual scenes based solely on the similarities and differences between them, with no labels or other information.

A pair of studies from researchers at the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center at MIT offers new evidence supporting this hypothesis. The researchers found that when they trained models known as neural networks using a particular type of self-supervised learning, the resulting models generated activity patterns very similar to those seen in the brains of animals that were performing the same tasks as the models.

The findings suggest that these models are able to learn representations of the physical world that they can use to make accurate predictions about what will happen in that world, and that the mammalian brain may be using the same strategy, the researchers say.

“The theme of our work is that AI designed to help build better robots ends up also being a framework to better understand the brain more generally,” says Aran Nayebi, a postdoc in the ICoN Center. “We can’t say if it’s the whole brain yet, but across scales and disparate brain areas, our results seem to be suggestive of an organizing principle.”

Nayebi is the lead author of one of the studies, co-authored with Rishi Rajalingham, a former MIT postdoc now at Meta Reality Labs, and senior authors Mehrdad Jazayeri, an associate professor of brain and cognitive sciences and a member of the McGovern Institute for Brain Research; and Robert Yang, an assistant professor of brain and cognitive sciences and an associate member of the McGovern Institute. Ila Fiete, director of the ICoN Center, a professor of brain and cognitive sciences, and an associate member of the McGovern Institute, is the senior author of the other study, which was co-led by Mikail Khona, an MIT graduate student, and Rylan Schaeffer, a former senior research associate at MIT.

Both studies will be presented at the 2023 Conference on Neural Information Processing Systems (NeurIPS) in December.

Modeling the physical world

Early models of computer vision mainly relied on supervised learning. Using this approach, models are trained to classify images that are each labeled with a name — cat, car, etc. The resulting models work well, but this type of training requires a great deal of human-labeled data.

To create a more efficient alternative, in recent years researchers have turned to models built through a technique known as contrastive self-supervised learning. This type of learning allows an algorithm to learn to classify objects based on how similar they are to each other, with no external labels provided.

“This is a very powerful method because you can now leverage very large modern data sets, especially videos, and really unlock their potential,” Nayebi says. “A lot of the modern AI that you see now, especially in the last couple years with ChatGPT and GPT-4, is a result of training a self-supervised objective function on a large-scale dataset to obtain a very flexible representation.”

These types of models, also called neural networks, consist of thousands or millions of processing units connected to each other. Each node has connections of varying strengths to other nodes in the network. As the network analyzes huge amounts of data, the strengths of those connections change as the network learns to perform the desired task.

As the model performs a particular task, the activity patterns of different units within the network can be measured. Each unit’s activity can be represented as a firing pattern, similar to the firing patterns of neurons in the brain. Previous work from Nayebi and others has shown that self-supervised models of vision generate activity similar to that seen in the visual processing system of mammalian brains.

In both of the new NeurIPS studies, the researchers set out to explore whether self-supervised computational models of other cognitive functions might also show similarities to the mammalian brain. In the study led by Nayebi, the researchers trained self-supervised models to predict the future state of their environment across hundreds of thousands of naturalistic videos depicting everyday scenarios.

“For the last decade or so, the dominant method to build neural network models in cognitive neuroscience is to train these networks on individual cognitive tasks. But models trained this way rarely generalize to other tasks,” Yang says. “Here we test whether we can build models for some aspect of cognition by first training on naturalistic data using self-supervised learning, then evaluating in lab settings.”

Once the model was trained, the researchers had it generalize to a task they call “Mental-Pong.” This is similar to the video game Pong, where a player moves a paddle to hit a ball traveling across the screen. In the Mental-Pong version, the ball disappears shortly before hitting the paddle, so the player has to estimate its trajectory in order to hit the ball.

The researchers found that the model was able to track the hidden ball’s trajectory with accuracy similar to that of neurons in the mammalian brain, which had been shown in a previous study by Rajalingham and Jazayeri to simulate its trajectory — a cognitive phenomenon known as “mental simulation.” Furthermore, the neural activation patterns seen within the model were similar to those seen in the brains of animals as they played the game — specifically, in a part of the brain called the dorsomedial frontal cortex. No other class of computational model has been able to match the biological data as closely as this one, the researchers say.

“There are many efforts in the machine learning community to create artificial intelligence,” Jazayeri says. “The relevance of these models to neurobiology hinges on their ability to additionally capture the inner workings of the brain. The fact that Aran’s model predicts neural data is really important as it suggests that we may be getting closer to building artificial systems that emulate natural intelligence.”

Navigating the world

The study led by Khona, Schaeffer, and Fiete focused on a type of specialized neurons known as grid cells. These cells, located in the entorhinal cortex, help animals to navigate, working together with place cells located in the hippocampus.

While place cells fire whenever an animal is in a specific location, grid cells fire only when the animal is at one of the vertices of a triangular lattice. Groups of grid cells create overlapping lattices of different sizes, which allows them to encode a large number of positions using a relatively small number of cells.

In recent studies, researchers have trained supervised neural networks to mimic grid cell function by predicting an animal’s next location based on its starting point and velocity, a task known as path integration. However, these models hinged on access to privileged information about absolute space at all times — information that the animal does not have.

Inspired by the striking coding properties of the multiperiodic grid-cell code for space, the MIT team trained a contrastive self-supervised model to both perform this same path integration task and represent space efficiently while doing so. For the training data, they used sequences of velocity inputs. The model learned to distinguish positions based on whether they were similar or different — nearby positions generated similar codes, but further positions generated more different codes.

“It’s similar to training models on images, where if two images are both heads of cats, their codes should be similar, but if one is the head of a cat and one is a truck, then you want their codes to repel,” Khona says. “We’re taking that same idea but applying it to spatial trajectories.”

Once the model was trained, the researchers found that the activation patterns of the nodes within the model formed several lattice patterns with different periods, very similar to those formed by grid cells in the brain.

“What excites me about this work is that it makes connections between mathematical work on the striking information-theoretic properties of the grid cell code and the computation of path integration,” Fiete says. “While the mathematical work was analytic — what properties does the grid cell code possess? — the approach of optimizing coding efficiency through self-supervised learning and obtaining grid-like tuning is synthetic: It shows what properties might be necessary and sufficient to explain why the brain has grid cells.”

The research was funded by the K. Lisa Yang ICoN Center, the National Institutes of Health, the Simons Foundation, the McKnight Foundation, the McGovern Institute, and the Helen Hay Whitney Foundation.

Study: Deep neural networks don’t see the world the way we do

Anne Trafton |

Human sensory systems are very good at recognizing objects that we see or words that we hear, even if the object is upside down or the word is spoken by a voice we’ve never heard.

Computational models known as deep neural networks can be trained to do the same thing, correctly identifying an image of a dog regardless of what color its fur is, or a word regardless of the pitch of the speaker’s voice. However, a new study from MIT neuroscientists has found that these models often also respond the same way to images or words that have no resemblance to the target.

When these neural networks were used to generate an image or a word that they responded to in the same way as a specific natural input, such as a picture of a bear, most of them generated images or sounds that were unrecognizable to human observers. This suggests that these models build up their own idiosyncratic “invariances” — meaning that they respond the same way to stimuli with very different features.

The findings offer a new way for researchers to evaluate how well these models mimic the organization of human sensory perception, says Josh McDermott, an associate professor of brain and cognitive sciences at MIT and a member of MIT’s McGovern Institute for Brain Research and Center for Brains, Minds, and Machines.

“This paper shows that you can use these models to derive unnatural signals that end up being very diagnostic of the representations in the model,” says McDermott, who is the senior author of the study. “This test should become part of a battery of tests that we as a field are using to evaluate models.”

Jenelle Feather PhD ’22, who is now a research fellow at the Flatiron Institute Center for Computational Neuroscience, is the lead author of the open-access paper, which appears today in Nature Neuroscience. Guillaume Leclerc, an MIT graduate student, and Aleksander Mądry, the Cadence Design Systems Professor of Computing at MIT, are also authors of the paper.

Different perceptions

In recent years, researchers have trained deep neural networks that can analyze millions of inputs (sounds or images) and learn common features that allow them to classify a target word or object roughly as accurately as humans do. These models are currently regarded as the leading models of biological sensory systems.

It is believed that when the human sensory system performs this kind of classification, it learns to disregard features that aren’t relevant to an object’s core identity, such as how much light is shining on it or what angle it’s being viewed from. This is known as invariance, meaning that objects are perceived to be the same even if they show differences in those less important features.

“Classically, the way that we have thought about sensory systems is that they build up invariances to all those sources of variation that different examples of the same thing can have,” Feather says. “An organism has to recognize that they’re the same thing even though they show up as very different sensory signals.”

The researchers wondered if deep neural networks that are trained to perform classification tasks might develop similar invariances. To try to answer that question, they used these models to generate stimuli that produce the same kind of response within the model as an example stimulus given to the model by the researchers.

They term these stimuli “model metamers,” reviving an idea from classical perception research whereby stimuli that are indistinguishable to a system can be used to diagnose its invariances. The concept of metamers was originally developed in the study of human perception to describe colors that look identical even though they are made up of different wavelengths of light.

To their surprise, the researchers found that most of the images and sounds produced in this way looked and sounded nothing like the examples that the models were originally given. Most of the images were a jumble of random-looking pixels, and the sounds resembled unintelligible noise. When researchers showed the images to human observers, in most cases the humans did not classify the images synthesized by the models in the same category as the original target example.

“They’re really not recognizable at all by humans. They don’t look or sound natural and they don’t have interpretable features that a person could use to classify an object or word,” Feather says.

The findings suggest that the models have somehow developed their own invariances that are different from those found in human perceptual systems. This causes the models to perceive pairs of stimuli as being the same despite their being wildly different to a human.

Idiosyncratic invariances

The researchers found the same effect across many different vision and auditory models. However, each of these models appeared to develop their own unique invariances. When metamers from one model were shown to another model, the metamers were just as unrecognizable to the second model as they were to human observers.

“The key inference from that is that these models seem to have what we call idiosyncratic invariances,” McDermott says. “They have learned to be invariant to these particular dimensions in the stimulus space, and it’s model-specific, so other models don’t have those same invariances.”

The researchers also found that they could induce a model’s metamers to be more recognizable to humans by using an approach called adversarial training. This approach was originally developed to combat another limitation of object recognition models, which is that introducing tiny, almost imperceptible changes to an image can cause the model to misrecognize it.

The researchers found that adversarial training, which involves including some of these slightly altered images in the training data, yielded models whose metamers were more recognizable to humans, though they were still not as recognizable as the original stimuli. This improvement appears to be independent of the training’s effect on the models’ ability to resist adversarial attacks, the researchers say.

“This particular form of training has a big effect, but we don’t really know why it has that effect,” Feather says. “That’s an area for future research.”

Analyzing the metamers produced by computational models could be a useful tool to help evaluate how closely a computational model mimics the underlying organization of human sensory perception systems, the researchers say.

“This is a behavioral test that you can run on a given model to see whether the invariances are shared between the model and human observers,” Feather says. “It could also be used to evaluate how idiosyncratic the invariances are within a given model, which could help uncover potential ways to improve our models in the future.”

The research was funded by the National Science Foundation, the National Institutes of Health, a Department of Energy Computational Science Graduate Fellowship, and a Friends of the McGovern Institute Fellowship.

Study decodes surprising approach mice take in learning

by David Orenstein |

Neuroscience discoveries ranging from the nature of memory to treatments for disease have depended on reading the minds of mice, so researchers need to truly understand what the rodents’ behavior is telling them during experiments. In a new study that examines learning from reward, MIT researchers deciphered some initially mystifying mouse behavior, yielding new ideas about how mice think and a mathematical tool to aid future research.

The task the mice were supposed to master is simple: Turn a wheel left or right to get a reward and then recognize when the reward direction switches. When neurotypical people play such “reversal learning” games they quickly infer the optimal approach: stick with the direction that works until it doesn’t and then switch right away. Notably, people with schizophrenia struggle with the task. In the new study in PLOS Computational Biology, mice surprised scientists by showing that while they were capable of learning the “win-stay, lose-shift” strategy, they nonetheless refused to fully adopt it.

“It is not that mice cannot form an inference-based model of this environment—they can,” said corresponding author Mriganka Sur, Newton Professor in The Picower Institute for Learning and Memory and MIT’s Department of Brain and Cognitive Sciences (BCS). “The surprising thing is that they don’t persist with it. Even in a single block of the game where you know the reward is 100 percent on one side, every so often they will try the other side.”

While the mouse motif of departing from the optimal strategy could be due to a failure to hold it in memory, said lead author and Sur Lab graduate student Nhat Le, another possibility is that mice don’t commit to the “win-stay, lose-shift” approach because they don’t trust that their circumstances will remain stable or predictable. Instead, they might deviate from the optimal regime to test whether the rules have changed. Natural settings, after all, are rarely stable or predictable.

“I’d like to think mice are smarter than we give them credit for,” Le said.

But regardless of which reason may cause the mice to mix strategies, added co-senior author Mehrdad Jazayeri, Associate Professor in BCS and the McGovern Institute for Brain Research, it is important for researchers to recognize that they do and to be able to tell when and how they are choosing one strategy or another.

“This study highlights the fact that, unlike the accepted wisdom, mice doing lab tasks do not necessarily adopt a stationary strategy and it offers a computationally rigorous approach to detect and quantify such non-stationarities,” he said. “This ability is important because when researchers record the neural activity, their interpretation of the underlying algorithms and mechanisms may be invalid when they do not take the animals’ shifting strategies into account.”

Tracking thinking

The research team, which also includes co-author Murat Yildirim, a former Sur lab postdoc who is now an assistant professor at the Cleveland Clinic Lerner Research Institute, initially expected that the mice might adopt one strategy or the other. They simulated the results they’d expect to see if the mice either adopted the optimal strategy of inferring a rule about the task, or more randomly surveying whether left or right turns were being rewarded. Mouse behavior on the task, even after days, varied widely but it never resembled the results simulated by just one strategy.

To differing, individual extents, mouse performance on the task reflected variance along three parameters: how quickly they switched directions after the rule switched, how long it took them to transition to the new direction, and how loyal they remained to the new direction. Across 21 mice, the raw data represented a surprising diversity of outcomes on a task that neurotypical humans uniformly optimize. But the mice clearly weren’t helpless. Their average performance significantly improved over time, even though it plateaued below the optimal level.

In the task, the rewarded side switched every 15-25 turns. The team realized the mice were using more than one strategy in each such “block” of the game, rather than just inferring the simple rule and optimizing based on that inference. To disentangle when the mice were employing that strategy or another, the team harnessed an analytical framework called a Hidden Markov Model (HMM), which can computationally tease out when one unseen state is producing a result vs. another unseen state. Le likens it to what a judge on a cooking show might do: inferring which chef contestant made which version of a dish based on patterns in each plate of food before them.

Before the team could use an HMM to decipher their mouse performance results, however, they had to adapt it. A typical HMM might apply to individual mouse choices, but here the team modified it to explain choice transitions over the course of whole blocks. They dubbed their modified model the blockHMM. Computational simulations of task performance using the blockHMM showed that the algorithm is able to infer the true hidden states of an artificial agent. The authors then used this technique to show the mice were persistently blending multiple strategies, achieving varied levels of performance.

“We verified that each animal executes a mixture of behavior from multiple regimes instead of a behavior in a single domain,” Le and his co-authors wrote. “Indeed 17/21 mice used a combination of low, medium and high-performance behavior modes.”

Further analysis revealed that the strategies afoot were indeed the “correct” rule inference strategy and a more exploratory strategy consistent with randomly testing options to get turn-by-turn feedback.

Now that the researchers have decoded the peculiar approach mice take to reversal learning, they are planning to look more deeply into the brain to understand which brain regions and circuits are involved. By watching brain cell activity during the task, they hope to discern what underlies the decisions the mice make to switch strategies.

By examining reversal learning circuits in detail, Sur said, it’s possible the team will gain insights that could help explain why people with schizophrenia show diminished performance on reversal learning tasks. Sur added that some people with autism spectrum disorders also persist with newly unrewarded behaviors longer than neurotypical people, so his lab will also have that phenomenon in mind as they investigate.

Yildirim, too, is interested in examining potential clinical connections.

“This reversal learning paradigm fascinates me since I want to use it in my lab with various preclinical models of neurological disorders,” he said. “The next step for us is to determine the brain mechanisms underlying these differences in behavioral strategies and whether we can manipulate these strategies.”

Funding for the study came from The National Institutes of Health, the Army Research Office, a Paul and Lilah Newton Brain Science Research Award, the Massachusetts Life Sciences Initiative, The Picower Institute for Learning and Memory and The JPB Foundation.

When computer vision works more like a brain, it sees more like people do

by Jennifer Michalowski |

From cameras to self-driving cars, many of today’s technologies depend on artificial intelligence (AI) to extract meaning from visual information.  Today’s AI technology has artificial neural networks at its core, and most of the time we can trust these AI computer vision systems to see things the way we do — but sometimes they falter. According to MIT and IBM Research scientists, one way to improve computer vision is to instruct the artificial neural networks that they rely on to deliberately mimic the way the brain’s biological neural network processes visual images.

Researchers led by James DiCarlo, the director of MIT’s Quest for Intelligence and member of the MIT-IBM Watson AI Lab, have made a computer vision model more robust by training it to work like a part of the brain that humans and other primates rely on for object recognition. This May, at the International Conference on Learning Representations (ICLR), the team reported that when they trained an artificial neural network using neural activity patterns in the brain’s inferior temporal (IT) cortex, the artificial neural network was more robustly able to identify objects in images than a model that lacked that neural training. And the model’s interpretations of images more closely matched what humans saw, even when images included minor distortions that made the task more difficult.

Comparing neural circuits

Portrait of Professor DiCarlo
McGovern Investigator and Director of MIT Quest for Intelligence, James DiCarlo. Photo: Justin Knight

Many of the artificial neural networks used for computer vision already resemble the multi-layered brain circuits that process visual information in humans and other primates. Like the brain, they use neuron-like units that work together to process information. As they are trained for a particular task, these layered components collectively and progressively process the visual information to complete the task — determining for example, that an image depicts a bear or a car or a tree.

DiCarlo and others previously found that when such deep-learning computer vision systems establish efficient ways to solve visual problems, they end up with artificial circuits that work similarly to the neural circuits that process visual information in our own brains. That is, they turn out to be surprisingly good scientific models of the neural mechanisms underlying primate and human vision.

That resemblance is helping neuroscientists deepen their understanding of the brain. By demonstrating ways visual information can be processed to make sense of images, computational models suggest hypotheses about how the brain might accomplish the same task. As developers continue to refine computer vision models, neuroscientists have found new ideas to explore in their own work.

“As vision systems get better at performing in the real world, some of them turn out to be more human-like in their internal processing. That’s useful from an understanding biology point of view,” says DiCarlo, who is also a professor of brain and cognitive sciences and an investigator at the McGovern Institute.

Engineering more brain-like AI

While their potential is promising, computer vision systems are not yet perfect models of human vision. DiCarlo suspected one way to improve computer vision may be to incorporate specific brain-like features into these models.

To test this idea, he and his collaborators built a computer vision model using neural data previously collected from vision-processing neurons in the monkey IT cortex — a key part of the primate ventral visual pathway involved in the recognition of objects — while the animals viewed various images. More specifically, Joel Dapello, a Harvard graduate student and former MIT-IBM Watson AI Lab intern, and Kohitij Kar, Assistant Professor, Canada Research Chair (Visual Neuroscience) at York University and visiting scientist at MIT, in collaboration with David Cox, IBM Research’s VP for AI Models and IBM director of the MIT-IBM Watson AI Lab, and other researchers at IBM Research and MIT, asked an artificial neural network to emulate the behavior of these primate vision-processing neurons while the network learned to identify objects in a standard computer vision task.

“In effect, we said to the network, ‘please solve this standard computer vision task, but please also make the function of one of your inside simulated “neural” layers be as similar as possible to the function of the corresponding biological neural layer,’” DiCarlo explains. “We asked it to do both of those things as best it could.” This forced the artificial neural circuits to find a different way to process visual information than the standard, computer vision approach, he says.

After training the artificial model with biological data, DiCarlo’s team compared its activity to a similarly-sized neural network model trained without neural data, using the standard approach for computer vision. They found that the new, biologically-informed model IT layer was – as instructed — a better match for IT neural data.  That is, for every image tested, the population of artificial IT neurons in the model responded more similarly to the corresponding population of biological IT neurons.

“Everybody gets something out of the exciting virtuous cycle between natural/biological intelligence and artificial intelligence,” DiCarlo says.

The researchers also found that the model IT was also a better match to IT neural data collected from another monkey, even though the model had never seen data from that animal, and even when that comparison was evaluated on that monkey’s IT responses to new images. This indicated that the team’s new, “neurally-aligned” computer model may be an improved model of the neurobiological function of the primate IT cortex — an interesting finding, given that it was previously unknown whether the amount of neural data that can be currently collected from the primate visual system is capable of directly guiding model development.

With their new computer model in hand, the team asked whether the “IT neural alignment” procedure also leads to any changes in the overall behavioral performance of the model. Indeed, they found that the neurally-aligned model was more human-like in its behavior — it tended to succeed in correctly categorizing objects in images for which humans also succeed, and it tended to fail when humans also fail.

Adversarial attacks

The team also found that the neurally-aligned model was more resistant to “adversarial attacks” that developers use to test computer vision and AI systems.  In computer vision, adversarial attacks introduce small distortions into images that are meant to mislead an artificial neural network.

“Say that you have an image that the model identifies as a cat. Because you have the knowledge of the internal workings of the model, you can then design very small changes in the image so that the model suddenly thinks it’s no longer a cat,” DiCarlo explains.

These minor distortions don’t typically fool humans, but computer vision models struggle with these alterations. A person who looks at the subtly distorted cat still reliably and robustly reports that it’s a cat. But standard computer vision models are more likely to mistake the cat for a dog, or even a tree.

“There must be some internal differences in the way our brains process images that lead to our vision being more resistant to those kinds of attacks,” DiCarlo says. And indeed, the team found that when they made their model more neurally-aligned, it became more robust, correctly identifying more images in the face of adversarial attacks.  The model could still be fooled by stronger “attacks,” but so can people, DiCarlo says. His team is now exploring the limits of adversarial robustness in humans.

A few years ago, DiCarlo’s team found they could also improve a model’s resistance to adversarial attacks by designing the first layer of the artificial network to emulate the early visual processing layer in the brain. One key next step is to combine such approaches — making new models that are simultaneously neurally-aligned at multiple visual processing layers.

The new work is further evidence that an exchange of ideas between neuroscience and computer science can drive progress in both fields. “Everybody gets something out of the exciting virtuous cycle between natural/biological intelligence and artificial intelligence,” DiCarlo says. “In this case, computer vision and AI researchers get new ways to achieve robustness and neuroscientists and cognitive scientists get more accurate mechanistic models of human vision.”

This work was supported by the MIT-IBM Watson AI Lab, Semiconductor Research Corporation, DARPA, the Massachusetts Institute of Technology Shoemaker Fellowship, Office of Naval Research, the Simons Foundation, and Canada Research Chair Program.