Second round of seed grants awarded to MIT scholars studying the impact and applications of generative AI

Last summer, MIT President Sally Kornbluth and Provost Cynthia Barnhart issued a call for papers to “articulate effective roadmaps, policy recommendations, and calls for action across the broad domain of generative AI.” The response to the call far exceeded expectations with 75 proposals submitted. Of those, 27 proposals were selected for seed funding.

In light of this enthusiastic response, Kornbluth and Barnhart announced a second call for proposals this fall.

“The groundswell of interest and the caliber of the ideas overall made clear that a second round was in order,” they said in their email to MIT’s research community this fall. This second call for proposals resulted in 53 submissions.

Following the second call, the faculty committee from the first round considered the proposals and selected 16 proposals to receive exploratory funding. Co-authored by interdisciplinary teams of faculty and researchers affiliated with all five of the Institute’s schools and the MIT Schwarzman College of Computing, the proposals offer insights and perspectives on the potential impact and applications of generative AI across a broad range of topics and disciplines.

Each selected research group will receive between $50,000 and $70,000 to create 10-page impact papers. Those papers will be shared widely via a publication venue managed and hosted by the MIT Press under the auspices of the MIT Open Publishing Services program.

As with the first round of papers, Thomas Tull, a member of the MIT School of Engineering Dean’s Advisory Council and a former innovation scholar at the School of Engineering, contributed funding to support the effort.

The selected papers are:

“A Road-map for End-to-end Privacy and Verifiability in Generative AI,” led by Alex Pentland, Srini Devadas, Lalana Kagal, and Vinod Vaikuntanathan;
“A Virtuous Cycle: Generative AI and Discovery in the Physical Sciences,” led by Philip Harris and Phiala Shanahan;
“Artificial Cambrian Intelligence: Generating New Forms of Visual Intelligence,” led by Ramesh Raskar and Tomaso A. Poggio;
“Artificial Fictions and the Value of AI-Generated Art,” led by Justin Khoo;
“GenAI for Improving Human-to-human Interactions with a Focus on Negotiations,” led by Lawrence Susskind;
“Generative AI as a New Applications Platform and Ecosystem,” led by Michael Cusumano;
“Generative AI for Cities: A Civic Engagement Playbook,” led by Sarah Williams, Sara Beery, and Eden Medina;
“Generative AI for Textile Engineering: Advanced Materials from Heritage Lace Craft,” led by Svetlana V. Boriskina;
“Generative AI Impact for Biomedical Innovation and Drug Discovery,” led by Manolis Kellis, Brad Pentelute, and Marinka Zitnik;
“Impact of Generative AI on the Creative Economy,” led by Ashia Wilson and Dylan Hadfield-Menell;
“Redefining Virtuosity: The Role of Generative AI in Live Music Performances,” led by Joseph A. Paradiso and Eran Egozy;
“Reflection-based Learning with Generative AI,” led by Stefanie Mueller;
“Robust and Reliable Systems for Generative AI,” led by Shafi Goldwasser, Yael Kalai, and Vinod Vaikuntanathan;
“Supporting the Aging Population with Generative AI,” led by Pattie Maes;
“The Science of Language in the Era of Generative AI,” led by Danny Fox, Yoon Kim, and Roger Levy; and
“Visual Artists, Technological Shock, and Generative AI,” led by Caroline Jones and Huma Gupta.

MIT-derived algorithm helps forecast the frequency of extreme weather

To assess a community’s risk of extreme weather, policymakers rely first on global climate models that can be run decades, and even centuries, forward in time, but only at a coarse resolution. These models might be used to gauge, for instance, future climate conditions for the northeastern U.S., but not specifically for Boston.

To estimate Boston’s future risk of extreme weather such as flooding, policymakers can combine a coarse model’s large-scale predictions with a finer-resolution model, tuned to estimate how often Boston is likely to experience damaging floods as the climate warms. But this risk analysis is only as accurate as the predictions from that first, coarser climate model.

“If you get those wrong for large-scale environments, then you miss everything in terms of what extreme events will look like at smaller scales, such as over individual cities,” says Themistoklis Sapsis, the William I. Koch Professor and director of the Center for Ocean Engineering in MIT’s Department of Mechanical Engineering.

Sapsis and his colleagues have now developed a method to “correct” the predictions from coarse climate models. By combining machine learning with dynamical systems theory, the team’s approach “nudges” a climate model’s simulations into more realistic patterns over large scales. When paired with smaller-scale models to predict specific weather events such as tropical cyclones or floods, the team’s approach produced more accurate predictions for how often specific locations will experience those events over the next few decades, compared to predictions made without the correction scheme.

Sapsis says the new correction scheme is general in form and can be applied to any global climate model. Once corrected, the models can help to determine where and how often extreme weather will strike as global temperatures rise over the coming years. 

“Climate change will have an effect on every aspect of human life, and every type of life on the planet, from biodiversity to food security to the economy,” Sapsis says. “If we have capabilities to know accurately how extreme weather will change, especially over specific locations, it can make a lot of difference in terms of preparation and doing the right engineering to come up with solutions. This is the method that can open the way to do that.”

The team’s results appear today in the Journal of Advances in Modeling Earth Systems. The study’s MIT co-authors include postdoc Benedikt Barthel Sorensen and Alexis-Tzianni Charalampopoulos SM ’19, PhD ’23, with Shixuan Zhang, Bryce Harrop, and Ruby Leung of the Pacific Northwest National Laboratory in Washington state.

Over the hood

Today’s large-scale climate models simulate weather features such as the average temperature, humidity, and precipitation around the world, on a grid-by-grid basis. Running simulations of these models takes enormous computing power, and in order to simulate how weather features will interact and evolve over periods of decades or longer, models average out features every 100 kilometers or so.

“It’s a very heavy computation requiring supercomputers,” Sapsis notes. “But these models still do not resolve very important processes like clouds or storms, which occur over smaller scales of a kilometer or less.”

To improve the resolution of these coarse climate models, scientists typically have gone under the hood to try and fix a model’s underlying dynamical equations, which describe how phenomena in the atmosphere and oceans should physically interact.

“People have tried to dissect into climate model codes that have been developed over the last 20 to 30 years, which is a nightmare, because you can lose a lot of stability in your simulation,” Sapsis explains. “What we’re doing is a completely different approach, in that we’re not trying to correct the equations but instead correct the model’s output.”

The team’s new approach takes a model’s output, or simulation, and overlays an algorithm that nudges the simulation toward something that more closely represents real-world conditions. The algorithm is based on a machine-learning scheme that takes in data, such as past information for temperature and humidity around the world, and learns associations within the data that represent fundamental dynamics among weather features. The algorithm then uses these learned associations to correct a model’s predictions.

“What we’re doing is trying to correct dynamics, as in how an extreme weather feature, such as the windspeeds during a Hurricane Sandy event, will look like in the coarse model, versus in reality,” Sapsis says. “The method learns dynamics, and dynamics are universal. Having the correct dynamics eventually leads to correct statistics, for example, frequency of rare extreme events.”

Climate correction

As a first test of their new approach, the team used the machine-learning scheme to correct simulations produced by the Energy Exascale Earth System Model (E3SM), a climate model run by the U.S. Department of Energy, that simulates climate patterns around the world at a resolution of 110 kilometers. The researchers used eight years of past data for temperature, humidity, and wind speed to train their new algorithm, which learned dynamical associations between the measured weather features and the E3SM model. They then ran the climate model forward in time for about 36 years and applied the trained algorithm to the model’s simulations. They found that the corrected version produced climate patterns that more closely matched real-world observations from the last 36 years, not used for training.

“We’re not talking about huge differences in absolute terms,” Sapsis says. “An extreme event in the uncorrected simulation might be 105 degrees Fahrenheit, versus 115 degrees with our corrections. But for humans experiencing this, that is a big difference.”

When the team then paired the corrected coarse model with a specific, finer-resolution model of tropical cyclones, they found the approach accurately reproduced the frequency of extreme storms in specific locations around the world.

“We now have a coarse model that can get you the right frequency of events, for the present climate. It’s much more improved,” Sapsis says. “Once we correct the dynamics, this is a relevant correction, even when you have a different average global temperature, and it can be used for understanding how forest fires, flooding events, and heat waves will look in a future climate. Our ongoing work is focusing on analyzing future climate scenarios.”

“The results are particularly impressive as the method shows promising results on E3SM, a state-of-the-art climate model,” says Pedram Hassanzadeh, an associate professor who leads the Climate Extremes Theory and Data group at the University of Chicago and was not involved with the study. “It would be interesting to see what climate change projections this framework yields once future greenhouse-gas emission scenarios are incorporated.”

This work was supported, in part, by the U.S. Defense Advanced Research Projects Agency.

Large language models use a surprisingly simple mechanism to retrieve some stored knowledge

Large language models, such as those that power popular artificial intelligence chatbots like ChatGPT, are incredibly complex. Even though these models are being used as tools in many areas, such as customer support, code generation, and language translation, scientists still don’t fully grasp how they work.

In an effort to better understand what is going on under the hood, researchers at MIT and elsewhere studied the mechanisms at work when these enormous machine-learning models retrieve stored knowledge.

They found a surprising result: Large language models (LLMs) often use a very simple linear function to recover and decode stored facts. Moreover, the model uses the same decoding function for similar types of facts. Linear functions, equations with only two variables and no exponents, capture the straightforward, straight-line relationship between two variables.

The researchers showed that, by identifying linear functions for different facts, they can probe the model to see what it knows about new subjects, and where within the model that knowledge is stored.

Using a technique they developed to estimate these simple functions, the researchers found that even when a model answers a prompt incorrectly, it has often stored the correct information. In the future, scientists could use such an approach to find and correct falsehoods inside the model, which could reduce a model’s tendency to sometimes give incorrect or nonsensical answers.

“Even though these models are really complicated, nonlinear functions that are trained on lots of data and are very hard to understand, there are sometimes really simple mechanisms working inside them. This is one instance of that,” says Evan Hernandez, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper detailing these findings.

Hernandez wrote the paper with co-lead author Arnab Sharma, a computer science graduate student at Northeastern University; his advisor, Jacob Andreas, an associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); senior author David Bau, an assistant professor of computer science at Northeastern; and others at MIT, Harvard University, and the Israeli Institute of Technology. The research will be presented at the International Conference on Learning Representations.

Finding facts

Most large language models, also called transformer models, are neural networks. Loosely based on the human brain, neural networks contain billions of interconnected nodes, or neurons, that are grouped into many layers, and which encode and process data.

Much of the knowledge stored in a transformer can be represented as relations that connect subjects and objects. For instance, “Miles Davis plays the trumpet” is a relation that connects the subject, Miles Davis, to the object, trumpet.

As a transformer gains more knowledge, it stores additional facts about a certain subject across multiple layers. If a user asks about that subject, the model must decode the most relevant fact to respond to the query.

If someone prompts a transformer by saying “Miles Davis plays the. . .” the model should respond with “trumpet” and not “Illinois” (the state where Miles Davis was born).

“Somewhere in the network’s computation, there has to be a mechanism that goes and looks for the fact that Miles Davis plays the trumpet, and then pulls that information out and helps generate the next word. We wanted to understand what that mechanism was,” Hernandez says.

The researchers set up a series of experiments to probe LLMs, and found that, even though they are extremely complex, the models decode relational information using a simple linear function. Each function is specific to the type of fact being retrieved.

For example, the transformer would use one decoding function any time it wants to output the instrument a person plays and a different function each time it wants to output the state where a person was born.

The researchers developed a method to estimate these simple functions, and then computed functions for 47 different relations, such as “capital city of a country” and “lead singer of a band.”

While there could be an infinite number of possible relations, the researchers chose to study this specific subset because they are representative of the kinds of facts that can be written in this way.

They tested each function by changing the subject to see if it could recover the correct object information. For instance, the function for “capital city of a country” should retrieve Oslo if the subject is Norway and London if the subject is England.

Functions retrieved the correct information more than 60 percent of the time, showing that some information in a transformer is encoded and retrieved in this way.

“But not everything is linearly encoded. For some facts, even though the model knows them and will predict text that is consistent with these facts, we can’t find linear functions for them. This suggests that the model is doing something more intricate to store that information,” he says.

Visualizing a model’s knowledge

They also used the functions to determine what a model believes is true about different subjects.

In one experiment, they started with the prompt “Bill Bradley was a” and used the decoding functions for “plays sports” and “attended university” to see if the model knows that Sen. Bradley was a basketball player who attended Princeton.

“We can show that, even though the model may choose to focus on different information when it produces text, it does encode all that information,” Hernandez says.

They used this probing technique to produce what they call an “attribute lens,” a grid that visualizes where specific information about a particular relation is stored within the transformer’s many layers.

Attribute lenses can be generated automatically, providing a streamlined method to help researchers understand more about a model. This visualization tool could enable scientists and engineers to correct stored knowledge and help prevent an AI chatbot from giving false information.

In the future, Hernandez and his collaborators want to better understand what happens in cases where facts are not stored linearly. They would also like to run experiments with larger models, as well as study the precision of linear decoding functions.

“This is an exciting work that reveals a missing piece in our understanding of how large language models recall factual knowledge during inference. Previous work showed that LLMs build information-rich representations of given subjects, from which specific attributes are being extracted during inference. This work shows that the complex nonlinear computation of LLMs for attribute extraction can be well-approximated with a simple linear function,” says Mor Geva Pipek, an assistant professor in the School of Computer Science at Tel Aviv University, who was not involved with this work.

This research was supported, in part, by Open Philanthropy, the Israeli Science Foundation, and an Azrieli Foundation Early Career Faculty Fellowship.

Engineering household robots to have a little common sense

From wiping up spills to serving up food, robots are being taught to carry out increasingly complicated household tasks. Many such home-bot trainees are learning through imitation; they are programmed to copy the motions that a human physically guides them through.

It turns out that robots are excellent mimics. But unless engineers also program them to adjust to every possible bump and nudge, robots don’t necessarily know how to handle these situations, short of starting their task from the top.

Now MIT engineers are aiming to give robots a bit of common sense when faced with situations that push them off their trained path. They’ve developed a method that connects robot motion data with the “common sense knowledge” of large language models, or LLMs.

Their approach enables a robot to logically parse many given household task into subtasks, and to physically adjust to disruptions within a subtask so that the robot can move on without having to go back and start a task from scratch — and without engineers having to explicitly program fixes for every possible failure along the way.   

“Imitation learning is a mainstream approach enabling household robots. But if a robot is blindly mimicking a human’s motion trajectories, tiny errors can accumulate and eventually derail the rest of the execution,” says Yanwei Wang, a graduate student in MIT’s Department of Electrical Engineering and Computer Science (EECS). “With our method, a robot can self-correct execution errors and improve overall task success.”

Wang and his colleagues detail their new approach in a study they will present at the International Conference on Learning Representations (ICLR) in May. The study’s co-authors include EECS graduate students Tsun-Hsuan Wang and Jiayuan Mao, Michael Hagenow, a postdoc in MIT’s Department of Aeronautics and Astronautics (AeroAstro), and Julie Shah, the H.N. Slater Professor in Aeronautics and Astronautics at MIT.

Language task

The researchers illustrate their new approach with a simple chore: scooping marbles from one bowl and pouring them into another. To accomplish this task, engineers would typically move a robot through the motions of scooping and pouring — all in one fluid trajectory. They might do this multiple times, to give the robot a number of human demonstrations to mimic.

“But the human demonstration is one long, continuous trajectory,” Wang says.

The team realized that, while a human might demonstrate a single task in one go, that task depends on a sequence of subtasks, or trajectories. For instance, the robot has to first reach into a bowl before it can scoop, and it must scoop up marbles before moving to the empty bowl, and so forth. If a robot is pushed or nudged to make a mistake during any of these subtasks, its only recourse is to stop and start from the beginning, unless engineers were to explicitly label each subtask and program or collect new demonstrations for the robot to recover from the said failure, to enable a robot to self-correct in the moment.

“That level of planning is very tedious,” Wang says.

Instead, he and his colleagues found some of this work could be done automatically by LLMs. These deep learning models process immense libraries of text, which they use to establish connections between words, sentences, and paragraphs. Through these connections, an LLM can then generate new sentences based on what it has learned about the kind of word that is likely to follow the last.

For their part, the researchers found that in addition to sentences and paragraphs, an LLM can be prompted to produce a logical list of subtasks that would be involved in a given task. For instance, if queried to list the actions involved in scooping marbles from one bowl into another, an LLM might produce a sequence of verbs such as “reach,” “scoop,” “transport,” and “pour.”

“LLMs have a way to tell you how to do each step of a task, in natural language. A human’s continuous demonstration is the embodiment of those steps, in physical space,” Wang says. “And we wanted to connect the two, so that a robot would automatically know what stage it is in a task, and be able to replan and recover on its own.”

Mapping marbles

For their new approach, the team developed an algorithm to automatically connect an LLM’s natural language label for a particular subtask with a robot’s position in physical space or an image that encodes the robot state. Mapping a robot’s physical coordinates, or an image of the robot state, to a natural language label is known as “grounding.” The team’s new algorithm is designed to learn a grounding “classifier,” meaning that it learns to automatically identify what semantic subtask a robot is in — for example, “reach” versus “scoop” — given its physical coordinates or an image view.

“The grounding classifier facilitates this dialogue between what the robot is doing in the physical space and what the LLM knows about the subtasks, and the constraints you have to pay attention to within each subtask,” Wang explains.

The team demonstrated the approach in experiments with a robotic arm that they trained on a marble-scooping task. Experimenters trained the robot by physically guiding it through the task of first reaching into a bowl, scooping up marbles, transporting them over an empty bowl, and pouring them in. After a few demonstrations, the team then used a pretrained LLM and asked the model to list the steps involved in scooping marbles from one bowl to another. The researchers then used their new algorithm to connect the LLM’s defined subtasks with the robot’s motion trajectory data. The algorithm automatically learned to map the robot’s physical coordinates in the trajectories and the corresponding image view to a given subtask.

The team then let the robot carry out the scooping task on its own, using the newly learned grounding classifiers. As the robot moved through the steps of the task, the experimenters pushed and nudged the bot off its path, and knocked marbles off its spoon at various points. Rather than stop and start from the beginning again, or continue blindly with no marbles on its spoon, the bot was able to self-correct, and completed each subtask before moving on to the next. (For instance, it would make sure that it successfully scooped marbles before transporting them to the empty bowl.)

“With our method, when the robot is making mistakes, we don’t need to ask humans to program or give extra demonstrations of how to recover from failures,” Wang says. “That’s super exciting because there’s a huge effort now toward training household robots with data collected on teleoperation systems. Our algorithm can now convert that training data into robust robot behavior that can do complex tasks, despite external perturbations.”

AI generates high-quality images 30 times faster in a single step

In our current age of artificial intelligence, computers can generate their own “art” by way of diffusion models, iteratively adding structure to a noisy initial state until a clear image or video emerges. Diffusion models have suddenly grabbed a seat at everyone’s table: Enter a few words and experience instantaneous, dopamine-spiking dreamscapes at the intersection of reality and fantasy. Behind the scenes, it involves a complex, time-intensive process requiring numerous iterations for the algorithm to perfect the image.

MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have introduced a new framework that simplifies the multi-step process of traditional diffusion models into a single step, addressing previous limitations. This is done through a type of teacher-student model: teaching a new computer model to mimic the behavior of more complicated, original models that generate images. The approach, known as distribution matching distillation (DMD), retains the quality of the generated images and allows for much faster generation. 

“Our work is a novel method that accelerates current diffusion models such as Stable Diffusion and DALLE-3 by 30 times,” says Tianwei Yin, an MIT PhD student in electrical engineering and computer science, CSAIL affiliate, and the lead researcher on the DMD framework. “This advancement not only significantly reduces computational time but also retains, if not surpasses, the quality of the generated visual content. Theoretically, the approach marries the principles of generative adversarial networks (GANs) with those of diffusion models, achieving visual content generation in a single step — a stark contrast to the hundred steps of iterative refinement required by current diffusion models. It could potentially be a new generative modeling method that excels in speed and quality.”

This single-step diffusion model could enhance design tools, enabling quicker content creation and potentially supporting advancements in drug discovery and 3D modeling, where promptness and efficacy are key.

Distribution dreams

DMD cleverly has two components. First, it uses a regression loss, which anchors the mapping to ensure a coarse organization of the space of images to make training more stable. Next, it uses a distribution matching loss, which ensures that the probability to generate a given image with the student model corresponds to its real-world occurrence frequency. To do this, it leverages two diffusion models that act as guides, helping the system understand the difference between real and generated images and making training the speedy one-step generator possible.

The system achieves faster generation by training a new network to minimize the distribution divergence between its generated images and those from the training dataset used by traditional diffusion models. “Our key insight is to approximate gradients that guide the improvement of the new model using two diffusion models,” says Yin. “In this way, we distill the knowledge of the original, more complex model into the simpler, faster one, while bypassing the notorious instability and mode collapse issues in GANs.” 

Yin and colleagues used pre-trained networks for the new student model, simplifying the process. By copying and fine-tuning parameters from the original models, the team achieved fast training convergence of the new model, which is capable of producing high-quality images with the same architectural foundation. “This enables combining with other system optimizations based on the original architecture to further accelerate the creation process,” adds Yin. 

When put to the test against the usual methods, using a wide range of benchmarks, DMD showed consistent performance. On the popular benchmark of generating images based on specific classes on ImageNet, DMD is the first one-step diffusion technique that churns out pictures pretty much on par with those from the original, more complex models, rocking a super-close Fréchet inception distance (FID) score of just 0.3, which is impressive, since FID is all about judging the quality and diversity of generated images. Furthermore, DMD excels in industrial-scale text-to-image generation and achieves state-of-the-art one-step generation performance. There’s still a slight quality gap when tackling trickier text-to-image applications, suggesting there’s a bit of room for improvement down the line. 

Additionally, the performance of the DMD-generated images is intrinsically linked to the capabilities of the teacher model used during the distillation process. In the current form, which uses Stable Diffusion v1.5 as the teacher model, the student inherits limitations such as rendering detailed depictions of text and small faces, suggesting that DMD-generated images could be further enhanced by more advanced teacher models. 

“Decreasing the number of iterations has been the Holy Grail in diffusion models since their inception,” says Fredo Durand, MIT professor of electrical engineering and computer science, CSAIL principal investigator, and a lead author on the paper. “We are very excited to finally enable single-step image generation, which will dramatically reduce compute costs and accelerate the process.” 

“Finally, a paper that successfully combines the versatility and high visual quality of diffusion models with the real-time performance of GANs,” says Alexei Efros, a professor of electrical engineering and computer science at the University of California at Berkeley who was not involved in this study. “I expect this work to open up fantastic possibilities for high-quality real-time visual editing.” 

Yin and Durand’s fellow authors are MIT electrical engineering and computer science professor and CSAIL principal investigator William T. Freeman, as well as Adobe research scientists Michaël Gharbi SM ’15, PhD ’18; Richard Zhang; Eli Shechtman; and Taesung Park. Their work was supported, in part, by U.S. National Science Foundation grants (including one for the Institute for Artificial Intelligence and Fundamental Interactions), the Singapore Defense Science and Technology Agency, and by funding from Gwangju Institute of Science and Technology and Amazon. Their work will be presented at the Conference on Computer Vision and Pattern Recognition in June.

New algorithm unlocks high-resolution insights for computer vision

Imagine yourself glancing at a busy street for a few moments, then trying to sketch the scene you saw from memory. Most people could draw the rough positions of the major objects like cars, people, and crosswalks, but almost no one can draw every detail with pixel-perfect accuracy. The same is true for most modern computer vision algorithms: They are fantastic at capturing high-level details of a scene, but they lose fine-grained details as they process information.

Now, MIT researchers have created a system called “FeatUp” that lets algorithms capture all of the high- and low-level details of a scene at the same time — almost like Lasik eye surgery for computer vision.

When computers learn to “see” from looking at images and videos, they build up “ideas” of what’s in a scene through something called “features.” To create these features, deep networks and visual foundation models break down images into a grid of tiny squares and process these squares as a group to determine what’s going on in a photo. Each tiny square is usually made up of anywhere from 16 to 32 pixels, so the resolution of these algorithms is dramatically smaller than the images they work with. In trying to summarize and understand photos, algorithms lose a ton of pixel clarity. 

The FeatUp algorithm can stop this loss of information and boost the resolution of any deep network without compromising on speed or quality. This allows researchers to quickly and easily improve the resolution of any new or existing algorithm. For example, imagine trying to interpret the predictions of a lung cancer detection algorithm with the goal of localizing the tumor. Applying FeatUp before interpreting the algorithm using a method like class activation maps (CAM) can yield a dramatically more detailed (16-32x) view of where the tumor might be located according to the model. 

FeatUp not only helps practitioners understand their models, but also can improve a panoply of different tasks like object detection, semantic segmentation (assigning labels to pixels in an image with object labels), and depth estimation. It achieves this by providing more accurate, high-resolution features, which are crucial for building vision applications ranging from autonomous driving to medical imaging.

“The essence of all computer vision lies in these deep, intelligent features that emerge from the depths of deep learning architectures. The big challenge of modern algorithms is that they reduce large images to  very small grids of ’smart‘ features, gaining intelligent insights but losing the finer details,” says Mark Hamilton, an MIT PhD student in electrical engineering and computer science, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) affiliate, and a co-lead author on a paper about the project. “FeatUp helps enable the best of both worlds: highly intelligent representations with the original image’s resolution. These high-resolution features significantly boost performance across a spectrum of computer vision tasks, from enhancing object detection and improving depth prediction to providing a deeper understanding of your network’s decision-making process through high-resolution analysis.” 

Resolution renaissance 

As these large AI models become more and more prevalent, there’s an increasing need to explain what they’re doing, what they’re looking at, and what they’re thinking. 

But how exactly can FeatUp discover these fine-grained details? Curiously, the secret lies in wiggling and jiggling images. 

In particular, FeatUp applies minor adjustments (like moving the image a few pixels to the left or right) and watches how an algorithm responds to these slight movements of the image. This results in hundreds of deep-feature maps that are all slightly different, which can be combined into a single crisp, high-resolution, set of deep features. “We imagine that some high-resolution features exist, and that when we wiggle them and blur them, they will match all of the original, lower-resolution features from the wiggled images. Our goal is to learn how to refine the low-resolution features into high-resolution features using this ‚game‘ that lets us know how well we are doing,” says Hamilton. This methodology is analogous to how algorithms can create a 3D model from multiple 2D images by ensuring that the predicted 3D object matches all of the 2D photos used to create it. In FeatUp’s case, they predict a high-resolution feature map that’s consistent with all of the low-resolution feature maps formed by jittering the original image.

The team notes that standard tools available in PyTorch were insufficient for their needs, and introduced a new type of deep network layer in their quest for a speedy and efficient solution. Their custom layer, a special joint bilateral upsampling operation, was over 100 times more efficient than a naive implementation in PyTorch. The team also showed this new layer could improve a wide variety of different algorithms including semantic segmentation and depth prediction. This layer improved the network’s ability to process and understand high-resolution details, giving any algorithm that used it a substantial performance boost. 

“Another application is something called small object retrieval, where our algorithm allows for precise localization of objects. For example, even in cluttered road scenes algorithms enriched with FeatUp can see tiny objects like traffic cones, reflectors, lights, and potholes where their low-resolution cousins fail. This demonstrates its capability to enhance coarse features into finely detailed signals,” says Stephanie Fu ’22, MNG ’23, a PhD student at the University of California at Berkeley and another co-lead author on the new FeatUp paper. “This is especially critical for time-sensitive tasks, like pinpointing a traffic sign on a cluttered expressway in a driverless car. This can not only improve the accuracy of such tasks by turning broad guesses into exact localizations, but might also make these systems more reliable, interpretable, and trustworthy.”

What next?

Regarding future aspirations, the team emphasizes FeatUp’s potential widespread adoption within the research community and beyond, akin to data augmentation practices. “The goal is to make this method a fundamental tool in deep learning, enriching models to perceive the world in greater detail without the computational inefficiency of traditional high-resolution processing,” says Fu.

“FeatUp represents a wonderful advance towards making visual representations really useful, by producing them at full image resolutions,” says Cornell University computer science professor Noah Snavely, who was not involved in the research. “Learned visual representations have become really good in the last few years, but they are almost always produced at very low resolution — you might put in a nice full-resolution photo, and get back a tiny, postage stamp-sized grid of features. That’s a problem if you want to use those features in applications that produce full-resolution outputs. FeatUp solves this problem in a creative way by combining classic ideas in super-resolution with modern learning approaches, leading to beautiful, high-resolution feature maps.”

“We hope this simple idea can have broad application. It provides high-resolution versions of image analytics that we’d thought before could only be low-resolution,” says senior author William T. Freeman, an MIT professor of electrical engineering and computer science professor and CSAIL member.

Lead authors Fu and Hamilton are accompanied by MIT PhD students Laura Brandt SM ’21 and Axel Feldmann SM ’21, as well as Zhoutong Zhang SM ’21, PhD ’22, all current or former affiliates of MIT CSAIL. Their research is supported, in part, by a National Science Foundation Graduate Research Fellowship, by the National Science Foundation and Office of the Director of National Intelligence, by the U.S. Air Force Research Laboratory, and by the U.S. Air Force Artificial Intelligence Accelerator. The group will present their work in May at the International Conference on Learning Representations.

Five MIT faculty members take on Cancer Grand Challenges

Cancer Grand Challenges recently announced five winning teams for 2024, which included five researchers from MIT: Michael Birnbaum, Regina Barzilay, Brandon DeKosky, Seychelle Vos, and Ömer Yilmaz. Each team is made up of interdisciplinary cancer researchers from across the globe and will be awarded $25 million over five years. 

Birnbaum, an associate professor in the Department of Biological Engineering, leads Team MATCHMAKERS and is joined by co-investigators Barzilay, the School of Engineering Distinguished Professor for AI and Health in the Department of Electrical Engineering and Computer Science and the AI faculty lead at the MIT Abdul Latif Jameel Clinic for Machine Learning in Health; and DeKosky, Phillip and Susan Ragon Career Development Professor of Chemical Engineering. All three are also affiliates of the Koch Institute for Integrative Cancer Research At MIT.

Team MATCHMAKERS will take advantage of recent advances in artificial intelligence to develop tools for personalized immunotherapies for cancer patients. Cancer immunotherapies, which recruit the patient’s own immune system against the disease, have transformed treatment for some cancers, but not for all types and not for all patients. 

T cells are one target for immunotherapies because of their central role in the immune response. These immune cells use receptors on their surface to recognize protein fragments called antigens on cancer cells. Once T cells attach to cancer antigens, they mark them for destruction by the immune system. However, T cell receptors are exceptionally diverse within one person’s immune system and from person to person, making it difficult to predict how any one cancer patient will respond to an immunotherapy.  

Team MATCHMAKERS will collect data on T cell receptors and the different antigens they target and build computer models to predict antigen recognition by different T cell receptors. The team’s overarching goal is to develop tools for predicting T cell recognition with simple clinical lab tests and designing antigen-specific immunotherapies. “If successful, what we learn on our team could help transform prediction of T cell receptor recognition from something that is only possible in a few sophisticated laboratories in the world, for a few people at a time, into a routine process,” says Birnbaum. 

“The MATCHMAKERS project draws on MIT’s long tradition of developing cutting-edge artificial intelligence tools for the benefit of society,” comments Ryan Schoenfeld, CEO of The Mark Foundation for Cancer Research. “Their approach to optimizing immunotherapy for cancer and many other diseases is exemplary of the type of interdisciplinary research The Mark Foundation prioritizes supporting.” In addition to The Mark Foundation, the MATCHMAKERS team is funded by Cancer Research UK and the U.S. National Cancer Institute.

Vos, the Robert A. Swanson (1969) Career Development Professor of Life Sciences and HHMI Freeman Hrabowksi Scholar in the Department of Biology, will be a co-investigator on Team KOODAC. The KOODAC team will develop new treatments for solid tumors in children, using protein degradation strategies to target previously “undruggable” drivers of cancers. KOODAC is funded by Cancer Research UK, France’s Institut National Du Cancer, and KiKa (Children Cancer Free Foundation) through Cancer Grand Challenges. 

As a co-investigator on team PROSPECT, Yilmaz, who is also a Koch Institute affiliate, will help address early-onset colorectal cancers, an emerging global problem among individuals younger than 50 years. The team seeks to elucidate pathways, risk factors, and molecules involved in the disease’s development. Team PROSPECT is supported by Cancer Research UK, the U.S. National Cancer Institute, the Bowelbabe Fund for Cancer Research UK, and France’s Institut National Du Cancer through Cancer Grand Challenges.  

3 Questions: What you need to know about audio deepfakes

Audio deepfakes have had a recent bout of bad press after an artificial intelligence-generated robocall purporting to be the voice of Joe Biden hit up New Hampshire residents, urging them not to cast ballots. Meanwhile, spear-phishers — phishing campaigns that target a specific person or group, especially using information known to be of interest to the target — go fishing for money, and actors aim to preserve their audio likeness.

What receives less press, however, are some of the uses of audio deepfakes that could actually benefit society. In this Q&A prepared for MIT News, postdoc Nauman Dawalatabad addresses concerns as well as potential upsides of the emerging tech. A fuller version of this interview can be seen at the video below.

Q: What ethical considerations justify the concealment of the source speaker’s identity in audio deepfakes, especially when this technology is used for creating innovative content?

A: The inquiry into why research is important in obscuring the identity of the source speaker, despite a large primary use of generative models for audio creation in entertainment, for example, does raise ethical considerations. Speech does not contain the information only about “who you are?” (identity) or “what you are speaking?” (content); it encapsulates a myriad of sensitive information including age, gender, accent, current health, and even cues about the upcoming future health conditions. For instance, our recent research paper on “Detecting Dementia from Long Neuropsychological Interviews” demonstrates the feasibility of detecting dementia from speech with considerably high accuracy. Moreover, there are multiple models that can detect gender, accent, age, and other information from speech with very high accuracy. There is a need for advancements in technology that safeguard against the inadvertent disclosure of such private data. The endeavor to anonymize the source speaker’s identity is not merely a technical challenge but a moral obligation to preserve individual privacy in the digital age.

Q: How can we effectively maneuver through the challenges posed by audio deepfakes in spear-phishing attacks, taking into account the associated risks, the development of countermeasures, and the advancement of detection techniques?

A: The deployment of audio deepfakes in spear-phishing attacks introduces multiple risks, including the propagation of misinformation and fake news, identity theft, privacy infringements, and the malicious alteration of content. The recent circulation of deceptive robocalls in Massachusetts exemplifies the detrimental impact of such technology. We also recently spoke with the spoke with The Boston Globe about this technology, and how easy and inexpensive it is to generate such deepfake audios.

Anyone without a significant technical background can easily generate such audio, with multiple available tools online. Such fake news from deepfake generators can disturb financial markets and even electoral outcomes. The theft of one’s voice to access voice-operated bank accounts and the unauthorized utilization of one’s vocal identity for financial gain are reminders of the urgent need for robust countermeasures. Further risks may include privacy violation, where an attacker can utilize the victim’s audio without their permission or consent. Further, attackers can also alter the content of the original audio, which can have a serious impact.

Two primary and prominent directions have emerged in designing systems to detect fake audio: artifact detection and liveness detection. When audio is generated by a generative model, the model introduces some artifact in the generated signal. Researchers design algorithms/models to detect these artifacts. However, there are some challenges with this approach due to increasing sophistication of audio deepfake generators. In the future, we may also see models with very small or almost no artifacts. Liveness detection, on the other hand, leverages the inherent qualities of natural speech, such as breathing patterns, intonations, or rhythms, which are challenging for AI models to replicate accurately. Some companies like Pindrop are developing such solutions for detecting audio fakes. 

Additionally, strategies like audio watermarking serve as proactive defenses, embedding encrypted identifiers within the original audio to trace its origin and deter tampering. Despite other potential vulnerabilities, such as the risk of replay attacks, ongoing research and development in this arena offer promising solutions to mitigate the threats posed by audio deepfakes.

Q: Despite their potential for misuse, what are some positive aspects and benefits of audio deepfake technology? How do you imagine the future relationship between AI and our experiences of audio perception will evolve?

A: Contrary to the predominant focus on the nefarious applications of audio deepfakes, the technology harbors immense potential for positive impact across various sectors. Beyond the realm of creativity, where voice conversion technologies enable unprecedented flexibility in entertainment and media, audio deepfakes hold transformative promise in health care and education sectors. My current ongoing work in the anonymization of patient and doctor voices in cognitive health-care interviews, for instance, facilitates the sharing of crucial medical data for research globally while ensuring privacy. Sharing this data among researchers fosters development in the areas of cognitive health care. The application of this technology in voice restoration represents a hope for individuals with speech impairments, for example, for ALS or dysarthric speech, enhancing communication abilities and quality of life.

I am very positive about the future impact of audio generative AI models. The future interplay between AI and audio perception is poised for groundbreaking advancements, particularly through the lens of psychoacoustics — the study of how humans perceive sounds. Innovations in augmented and virtual reality, exemplified by devices like the Apple Vision Pro and others, are pushing the boundaries of audio experiences towards unparalleled realism. Recently we have seen an exponential increase in the number of sophisticated models coming up almost every month. This rapid pace of research and development in this field promises not only to refine these technologies but also to expand their applications in ways that profoundly benefit society. Despite the inherent risks, the potential for audio generative AI models to revolutionize health care, entertainment, education, and beyond is a testament to the positive trajectory of this research field.

Exploring the cellular neighborhood

Cells rely on complex molecular machines composed of protein assemblies to perform essential functions such as energy production, gene expression, and protein synthesis. To better understand how these machines work, scientists capture snapshots of them by isolating proteins from cells and using various methods to determine their structures. However, isolating proteins from cells also removes them from the context of their native environment, including protein interaction partners and cellular location.

Recently, cryogenic electron tomography (cryo-ET) has emerged as a way to observe proteins in their native environment by imaging frozen cells at different angles to obtain three-dimensional structural information. This approach is exciting because it allows researchers to directly observe how and where proteins associate with each other, revealing the cellular neighborhood of those interactions within the cell.

With the technology available to image proteins in their native environment, MIT graduate student Barrett Powell wondered if he could take it one step further: What if molecular machines could be observed in action? In a paper published March 8 in Nature Methods, Powell describes the method he developed, called tomoDRGN, for modeling structural differences of proteins in cryo-ET data that arise from protein motions or proteins binding to different interaction partners. These variations are known as structural heterogeneity. 

Although Powell had joined the lab of MIT associate professor of biology Joey Davis as an experimental scientist, he recognized the potential impact of computational approaches in understanding structural heterogeneity within a cell. Previously, the Davis Lab developed a related methodology named cryoDRGN to understand structural heterogeneity in purified samples. As Powell and Davis saw cryo-ET rising in prominence in the field, Powell took on the challenge of re-imagining this framework to work in cells.

When solving structures with purified samples, each particle is imaged only once. By contrast, cryo-ET data is collected by imaging each particle more than 40 times from different angles. That meant tomoDRGN needed to be able to merge the information from more than 40 images, which was where the project hit a roadblock: the amount of data led to an information overload.

To address this, Powell successfully rebuilt the cryoDRGN model to prioritize only the highest-quality data. When imaging the same particle multiple times, radiation damage occurs. The images acquired earlier, therefore, tend to be of higher quality because the particles are less damaged.

“By excluding some of the lower-quality data, the results were actually better than using all of the data — and the computational performance was substantially faster,” Powell says.

Just as Powell was beginning work on testing his model, he had a stroke of luck: The authors of a groundbreaking new study that visualized, for the first time, ribosomes inside cells at near-atomic resolution, shared their raw data on the Electric Microscopy Public Image Archive (EMPIAR). This dataset was an exemplary test case for Powell, through which he demonstrated that tomoDRGN could uncover structural heterogeneity within cryo-ET data. 

According to Powell, one exciting result is what tomoDRGN found surrounding a subset of ribosomes in the EMPIAR dataset. Some of the ribosomal particles were associated with a bacterial cell membrane and engaged in a process called cotranslational translocation. This occurs when a protein is being simultaneously synthesized and transported across a membrane. Researchers can use this result to make new hypotheses about how the ribosome functions with other protein machinery integral to transporting proteins outside of the cell, now guided by a structure of the complex in its native environment. 

After seeing that tomoDRGN could resolve structural heterogeneity from a structurally diverse dataset, Powell was curious: How small of a population could tomoDRGN identify? For that test, he chose a protein named apoferritin, which is a commonly used benchmark for cryo-ET and is often treated as structurally homogeneous. Ferritin is a protein used for iron storage and is referred to as apoferritin when it lacks iron.

Surprisingly, in addition to the expected particles, tomoDRGN revealed a minor population of ferritin particles — with iron bound — making up just 2 percent of the dataset, that was not previously reported. This result further demonstrated tomoDRGN’s ability to identify structural states that occur so infrequently that they would be averaged out of a 3D reconstruction. 

Powell and other members of the Davis Lab are excited to see how tomoDRGN can be applied to further ribosomal studies and to other systems. Davis works on understanding how cells assemble, regulate, and degrade molecular machines, so the next steps include exploring ribosome biogenesis within cells in greater detail using this new tool.

“What are the possible states that we may be losing during purification?” Davis asks. “Perhaps more excitingly, we can look at how they localize within the cell and what partners and protein complexes they may be interacting with.”

Researchers enhance peripheral vision in AI models

Peripheral vision enables humans to see shapes that aren’t directly in our line of sight, albeit with less detail. This ability expands our field of vision and can be helpful in many situations, such as detecting a vehicle approaching our car from the side.

Unlike humans, AI does not have peripheral vision. Equipping computer vision models with this ability could help them detect approaching hazards more effectively or predict whether a human driver would notice an oncoming object.

Taking a step in this direction, MIT researchers developed an image dataset that allows them to simulate peripheral vision in machine learning models. They found that training models with this dataset improved the models’ ability to detect objects in the visual periphery, although the models still performed worse than humans.

Their results also revealed that, unlike with humans, neither the size of objects nor the amount of visual clutter in a scene had a strong impact on the AI’s performance.

“There is something fundamental going on here. We tested so many different models, and even when we train them, they get a little bit better but they are not quite like humans. So, the question is: What is missing in these models?” says Vasha DuTell, a postdoc and co-author of a paper detailing this study.

Answering that question may help researchers build machine learning models that can see the world more like humans do. In addition to improving driver safety, such models could be used to develop displays that are easier for people to view.

Plus, a deeper understanding of peripheral vision in AI models could help researchers better predict human behavior, adds lead author Anne Harrington MEng ’23.

“Modeling peripheral vision, if we can really capture the essence of what is represented in the periphery, can help us understand the features in a visual scene that make our eyes move to collect more information,” she explains.

Their co-authors include Mark Hamilton, an electrical engineering and computer science graduate student; Ayush Tewari, a postdoc; Simon Stent, research manager at the Toyota Research Institute; and senior authors William T. Freeman, the Thomas and Gerd Perkins Professor of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Ruth Rosenholtz, principal research scientist in the Department of Brain and Cognitive Sciences and a member of CSAIL. The research will be presented at the International Conference on Learning Representations.

“Any time you have a human interacting with a machine — a car, a robot, a user interface — it is hugely important to understand what the person can see. Peripheral vision plays a critical role in that understanding,” Rosenholtz says.

Simulating peripheral vision

Extend your arm in front of you and put your thumb up — the small area around your thumbnail is seen by your fovea, the small depression in the middle of your retina that provides the sharpest vision. Everything else you can see is in your visual periphery. Your visual cortex represents a scene with less detail and reliability as it moves farther from that sharp point of focus.

Many existing approaches to model peripheral vision in AI represent this deteriorating detail by blurring the edges of images, but the information loss that occurs in the optic nerve and visual cortex is far more complex.

For a more accurate approach, the MIT researchers started with a technique used to model peripheral vision in humans. Known as the texture tiling model, this method transforms images to represent a human’s visual information loss.  

They modified this model so it could transform images similarly, but in a more flexible way that doesn’t require knowing in advance where the person or AI will point their eyes.

“That let us faithfully model peripheral vision the same way it is being done in human vision research,” says Harrington.

The researchers used this modified technique to generate a huge dataset of transformed images that appear more textural in certain areas, to represent the loss of detail that occurs when a human looks further into the periphery.

Then they used the dataset to train several computer vision models and compared their performance with that of humans on an object detection task.

“We had to be very clever in how we set up the experiment so we could also test it in the machine learning models. We didn’t want to have to retrain the models on a toy task that they weren’t meant to be doing,” she says.

Peculiar performance

Humans and models were shown pairs of transformed images which were identical, except that one image had a target object located in the periphery. Then, each participant was asked to pick the image with the target object.

“One thing that really surprised us was how good people were at detecting objects in their periphery. We went through at least 10 different sets of images that were just too easy. We kept needing to use smaller and smaller objects,” Harrington adds.

The researchers found that training models from scratch with their dataset led to the greatest performance boosts, improving their ability to detect and recognize objects. Fine-tuning a model with their dataset, a process that involves tweaking a pretrained model so it can perform a new task, resulted in smaller performance gains.

But in every case, the machines weren’t as good as humans, and they were especially bad at detecting objects in the far periphery. Their performance also didn’t follow the same patterns as humans.

“That might suggest that the models aren’t using context in the same way as humans are to do these detection tasks. The strategy of the models might be different,” Harrington says.

The researchers plan to continue exploring these differences, with a goal of finding a model that can predict human performance in the visual periphery. This could enable AI systems that alert drivers to hazards they might not see, for instance. They also hope to inspire other researchers to conduct additional computer vision studies with their publicly available dataset.

“This work is important because it contributes to our understanding that human vision in the periphery should not be considered just impoverished vision due to limits in the number of photoreceptors we have, but rather, a representation that is optimized for us to perform tasks of real-world consequence,” says Justin Gardner, an associate professor in the Department of Psychology at Stanford University who was not involved with this work. “Moreover, the work shows that neural network models, despite their advancement in recent years, are unable to match human performance in this regard, which should lead to more AI research to learn from the neuroscience of human vision. This future research will be aided significantly by the database of images provided by the authors to mimic peripheral human vision.”

This work is supported, in part, by the Toyota Research Institute and the MIT CSAIL METEOR Fellowship.