Kategorie: KI Use Cases

Words, data, and algorithms combine,
An article about LLMs, so divine. 
A glimpse into a linguistic world, 
Where language machines are unfurled.

It was a natural inclination to task a large language model (LLM) like CHATGPT with creating a poem that delves into the topic of large language models, and subsequently utilize said poem as an introductory piece for this article.

So how exactly did said poem get all stitched together in a neat package, with rhyming words and little morsels of clever phrases? 

We went straight to the source: MIT assistant professor and CSAIL principal investigator Jacob Andreas, whose research focuses on advancing the field of natural language processing, in both developing cutting-edge machine learning models and exploring the potential of language as a means of enhancing other forms of artificial intelligence. This includes pioneering work in areas such as using natural language to teach robots, and leveraging language to enable computer vision systems to articulate the rationale behind their decision-making processes. We probed Andreas regarding the mechanics, implications, and future prospects of the technology at hand.

Q: Language is a rich ecosystem ripe with subtle nuances that humans use to communicate with one another — sarcasm, irony, and other forms of figurative language. There’s numerous ways to convey meaning beyond the literal. Is it possible for large language models to comprehend the intricacies of context? What does it mean for a model to achieve „in-context learning“? Moreover, how do multilingual transformers process variations and dialects of different languages beyond English? 

A: When we think about linguistic contexts, these models are capable of reasoning about much, much longer documents and chunks of text more broadly than really anything that we’ve known how to build before. But that’s only one kind of context. With humans, language production and comprehension takes place in a grounded context. For example, I know that I’m sitting at this table. There are objects that I can refer to, and the language models we have right now typically can’t see any of that when interacting with a human user. 

There’s a broader social context that informs a lot of our language use which these models are, at least not immediately, sensitive to or aware of. It’s not clear how to give them information about the social context in which their language generation and language modeling takes place. Another important thing is temporal context. We’re shooting this video at a particular moment in time when particular facts are true. The models that we have right now were trained on, again, a snapshot of the internet that stopped at a particular time — for most models that we have now, probably a couple of years ago — and they don’t know about anything that’s happened since then. They don’t even know at what moment in time they’re doing text generation. Figuring out how to provide all of those different kinds of contexts is also an interesting question.

Maybe one of the most surprising components here is this phenomenon called in-context learning. If I take a small ML [machine learning] dataset and feed it to the model, like a movie review and the star rating assigned to the movie by the critic, you give just a couple of examples of these things, language models generate the ability both to generate plausible sounding movie reviews but also to predict the star ratings. More generally, if I have a machine learning problem, I have my inputs and my outputs. As you give an input to the model, you give it one more input and ask it to predict the output, the models can often do this really well.

This is a super interesting, fundamentally different way of doing machine learning, where I have this one big general-purpose model into which I can insert lots of little machine learning datasets, and yet without having to train a new model at all, classifier or a generator or whatever specialized to my particular task. This is actually something we’ve been thinking a lot about in my group, and in some collaborations with colleagues at Google — trying to understand exactly how this in-context learning phenomenon actually comes about.

Q: We like to believe humans are (at least somewhat) in pursuit of what is objectively and morally known to be true. Large language models, perhaps with under-defined or yet-to-be-understood „moral compasses,“ aren’t beholden to the truth. Why do large language models tend to hallucinate facts, or confidently assert inaccuracies? Does that limit the usefulness for applications where factual accuracy is critical? Is there a leading theory on how we will solve this? 

A: It’s well-documented that these models hallucinate facts, that they’re not always reliable. Recently, I asked ChatGPT to describe some of our group’s research. It named five papers, four of which are not papers that actually exist, and one of which is a real paper that was written by a colleague of mine who lives in the United Kingdom, whom I’ve never co-authored with. Factuality is still a big problem. Even beyond that, things involving reasoning in a really general sense, things involving complicated computations, complicated inferences, still seem to be really difficult for these models. There might be even fundamental limitations of this transformer architecture, and I believe a lot more modeling work is needed to make things better.

Why it happens is still partly an open question, but possibly, just architecturally, there are reasons that it’s hard for these models to build coherent models of the world. They can do that a little bit. You can query them with factual questions, trivia questions, and they get them right most of the time, maybe even more often than your average human user off the street. But unlike your average human user, it’s really unclear whether there’s anything that lives inside this language model that corresponds to a belief about the state of the world. I think this is both for architectural reasons, that transformers don’t, obviously, have anywhere to put that belief, and training data, that these models are trained on the internet, which was authored by a bunch of different people at different moments who believe different things about the state of the world. Therefore, it’s difficult to expect models to represent those things coherently.

All that being said, I don’t think this is a fundamental limitation of neural language models or even more general language models in general, but something that’s true about today’s language models. We’re already seeing that models are approaching being able to build representations of facts, representations of the state of the world, and I think there’s room to improve further.

Q: The pace of progress from GPT-2 to GPT-3 to GPT-4 has been dizzying. What does the pace of the trajectory look like from here? Will it be exponential, or an S-curve that will diminish in progress in the near term? If so, are there limiting factors in terms of scale, compute, data, or architecture?

A: Certainly in the short term, the thing that I’m most scared about has to do with these truthfulness and coherence issues that I was mentioning before, that even the best models that we have today do generate incorrect facts. They generate code with bugs, and because of the way these models work, they do so in a way that’s particularly difficult for humans to spot because the model output has all the right surface statistics. When we think about code, it’s still an open question whether it’s actually less work for somebody to write a function by hand or to ask a language model to generate that function and then have the person go through and verify that the implementation of that function was actually correct.

There’s a little danger in rushing to deploy these tools right away, and that we’ll wind up in a world where everything’s a little bit worse, but where it’s actually very difficult for people to actually reliably check the outputs of these models. That being said, these are problems that can be overcome. The pace that things are moving at especially, there’s a lot of room to address these issues of factuality and coherence and correctness of generated code in the long term. These really are tools, tools that we can use to free ourselves up as a society from a lot of unpleasant tasks, chores, or drudge work that has been difficult to automate — and that’s something to be excited about.

In an effort to improve fairness or reduce backlogs, machine-learning models are sometimes designed to mimic human decision making, such as deciding whether social media posts violate toxic content policies.

But researchers from MIT and elsewhere have found that these models often do not replicate human decisions about rule violations. If models are not trained with the right data, they are likely to make different, often harsher judgements than humans would.

In this case, the “right” data are those that have been labeled by humans who were explicitly asked whether items defy a certain rule. Training involves showing a machine-learning model millions of examples of this “normative data” so it can learn a task.

But data used to train machine-learning models are typically labeled descriptively — meaning humans are asked to identify factual features, such as, say, the presence of fried food in a photo. If “descriptive data” are used to train models that judge rule violations, such as whether a meal violates a school policy that prohibits fried food, the models tend to over-predict rule violations.

This drop in accuracy could have serious implications in the real world. For instance, if a descriptive model is used to make decisions about whether an individual is likely to reoffend, the researchers’ findings suggest it may cast stricter judgements than a human would, which could lead to higher bail amounts or longer criminal sentences.

“I think most artificial intelligence/machine-learning researchers assume that the human judgements in data and labels are biased, but this result is saying something worse. These models are not even reproducing already-biased human judgments because the data they’re being trained on has a flaw: Humans would label the features of images and text differently if they knew those features would be used for a judgment. This has huge ramifications for machine learning systems in human processes,” says Marzyeh Ghassemi, an assistant professor and head of the Healthy ML Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL).

Ghassemi is senior author of a new paper detailing these findings, which was published today in Science Advances. Joining her on the paper are lead author Aparna Balagopalan, an electrical engineering and computer science graduate student; David Madras, a graduate student at the University of Toronto; David H. Yang, a former graduate student who is now co-founder of ML Estimation; Dylan Hadfield-Menell, an MIT assistant professor; and Gillian K. Hadfield, Schwartz Reisman Chair in Technology and Society and professor of law at the University of Toronto.

Labeling discrepancy

This study grew out of a different project that explored how a machine-learning model can justify its predictions. As they gathered data for that study, the researchers noticed that humans sometimes give different answers if they are asked to provide descriptive or normative labels about the same data.

To gather descriptive labels, researchers ask labelers to identify factual features — does this text contain obscene language? To gather normative labels, researchers give labelers a rule and ask if the data violates that rule — does this text violate the platform’s explicit language policy?

Surprised by this finding, the researchers launched a user study to dig deeper. They gathered four datasets to mimic different policies, such as a dataset of dog images that could be in violation of an apartment’s rule against aggressive breeds. Then they asked groups of participants to provide descriptive or normative labels.

In each case, the descriptive labelers were asked to indicate whether three factual features were present in the image or text, such as whether the dog appears aggressive. Their responses were then used to craft judgements. (If a user said a photo contained an aggressive dog, then the policy was violated.) The labelers did not know the pet policy. On the other hand, normative labelers were given the policy prohibiting aggressive dogs, and then asked whether it had been violated by each image, and why.

The researchers found that humans were significantly more likely to label an object as a violation in the descriptive setting. The disparity, which they computed using the absolute difference in labels on average, ranged from 8 percent on a dataset of images used to judge dress code violations to 20 percent for the dog images.

“While we didn’t explicitly test why this happens, one hypothesis is that maybe how people think about rule violations is different from how they think about descriptive data. Generally, normative decisions are more lenient,” Balagopalan says.

Yet data are usually gathered with descriptive labels to train a model for a particular machine-learning task. These data are often repurposed later to train different models that perform normative judgements, like rule violations.

Training troubles

To study the potential impacts of repurposing descriptive data, the researchers trained two models to judge rule violations using one of their four data settings. They trained one model using descriptive data and the other using normative data, and then compared their performance.

They found that if descriptive data are used to train a model, it will underperform a model trained to perform the same judgements using normative data. Specifically, the descriptive model is more likely to misclassify inputs by falsely predicting a rule violation. And the descriptive model’s accuracy was even lower when classifying objects that human labelers disagreed about.

“This shows that the data do really matter. It is important to match the training context to the deployment context if you are training models to detect if a rule has been violated,” Balagopalan says.

It can be very difficult for users to determine how data have been gathered; this information can be buried in the appendix of a research paper or not revealed by a private company, Ghassemi says.

Improving dataset transparency is one way this problem could be mitigated. If researchers know how data were gathered, then they know how those data should be used. Another possible strategy is to fine-tune a descriptively trained model on a small amount of normative data. This idea, known as transfer learning, is something the researchers want to explore in future work.

They also want to conduct a similar study with expert labelers, like doctors or lawyers, to see if it leads to the same label disparity.

“The way to fix this is to transparently acknowledge that if we want to reproduce human judgment, we must only use data that were collected in that setting. Otherwise, we are going to end up with systems that are going to have extremely harsh moderations, much harsher than what humans would do. Humans would see nuance or make another distinction, whereas these models don’t,” Ghassemi says.

This research was funded, in part, by the Schwartz Reisman Institute for Technology and Society, Microsoft Research, the Vector Institute, and a Canada Research Council Chain.

Citadel founder and CEO Ken Griffin had some free advice for an at-capacity crowd of MIT students at the Wong Auditorium during a campus visit in April. “If you find yourself in a career where you’re not learning,” he told them, “it’s time to change jobs. In this world, if you’re not learning, you can find yourself irrelevant in the blink of an eye.”

During a conversation with Bryan Landman ’11, senior quantitative research lead for Citadel’s Global Quantitative Strategies business, Griffin reflected on his career and offered predictions for the impact of technology on the finance sector. Citadel, which he launched in 1990, is now one of the world’s leading investment firms. Griffin also serves as non-executive chair of Citadel Securities, a market maker that is known as a key player in the modernization of markets and market structures.

“We are excited to hear Ken share his perspective on how technology continues to shape the future of finance, including the emerging trends of quantum computing and AI,” said David Schmittlein, the John C Head III Dean and professor of marketing at MIT Sloan School of Management, who kicked off the program. The presentation was jointly sponsored by MIT Sloan, the MIT Schwarzman College of Computing, the School of Engineering, MIT Career Advising and Professional Development, and Citadel Securities Campus Recruiting.

The future, in Griffin’s view, “is all about the application of engineering, software, and mathematics to markets. Successful entrepreneurs are those who have the tools to solve the unsolved problems of that moment in time.” He launched Citadel only one year after graduating from college. “History so far has been kind to the vision I had back in the late ’80s,” he said.

Griffin realized very early in his career “that you could use a personal computer and quantitative finance to price traded securities in a way that was much more advanced than you saw on your typical equity trading desk on Wall Street.” Both businesses, he told the audience, are ultimately driven by research. “That’s where we formulate the ideas, and trading is how we monetize that research.”

It’s also why Citadel and Citadel Securities employ several hundred software engineers. “We have a huge investment today in using modern technology to power our decision-making and trading,” said Griffin.

One example of Citadel’s application of technology and science is the firm’s hiring of a meteorological team to expand the weather analytics expertise within its commodities business. While power supply is relatively easy to map and analyze, predicting demand is much more difficult. Citadel’s weather team feeds forecast data obtained from supercomputers to its traders. “Wind and solar are huge commodities,” Griffin explained, noting that the days with highest demand in the power market are cloudy, cold days with no wind. When you can forecast those days better than the market as a whole, that’s where you can identify opportunities, he added.

Pros and cons of machine learning

Asking about the impact of new technology on their sector, Landman noted that both Citadel and Citadel Securities are already leveraging machine learning. “In the market-making business,” Griffin said, “you see a real application for machine learning because you have so much data to parametrize the models with. But when you get into longer time horizon problems, machine learning starts to break down.”

Griffin noted that the data obtained through machine learning is most helpful for investments with short time horizons, such as in its quantitative strategies business. “In our fundamental equities business,” he said, “machine learning is not as helpful as you would want because the underlying systems are not stationary.”

Griffin was emphatic that “there has been a moment in time where being a really good statistician or really understanding machine-learning models was sufficient to make money. That won’t be the case for much longer.” One of the guiding principles at Citadel, he and Landman agreed, was that machine learning and other methodologies should not be used blindly. Each analyst has to cite the underlying economic theory driving their argument on investment decisions. “If you understand the problem in a different way than people who are just using the statistical models,” he said, “you have a real chance for a competitive advantage.”

ChatGPT and a seismic shift

Asked if ChatGPT will change history, Griffin predicted that the rise of capabilities in large language models will transform a substantial number of white collar jobs. “With open AI for most routine commercial legal documents, ChatGPT will do a better job writing a lease than a young lawyer. This is the first time we are seeing traditionally white-collar jobs at risk due to technology, and that’s a sea change.”

Griffin urged MIT students to work with the smartest people they can find, as he did: “The magic of Citadel has been a testament to the idea that by surrounding yourself with bright, ambitious people, you can accomplish something special. I went to great lengths to hire the brightest people I could find and gave them responsibility and trust early in their careers.”

Even more critical to success is the willingness to advocate for oneself, Griffin said, using Gerald Beeson, Citadel’s chief operating officer, as an example. Beeson, who started as an intern at the firm, “consistently sought more responsibility and had the foresight to train his own successors.” Urging students to take ownership of their careers, Griffin advised: “Make it clear that you’re willing to take on more responsibility, and think about what the roadblocks will be.”

When microphones were handed to the audience, students inquired what changes Griffin would like to see in the hedge fund industry, how Citadel assesses the risk and reward of potential projects, and whether hedge funds should give back to the open source community. Asked about the role that Citadel — and its CEO — should play in “the wider society,” Griffin spoke enthusiastically of his belief in participatory democracy. “We need better people on both sides of the aisle,” he said. “I encourage all my colleagues to be politically active. It’s unfortunate when firms shut down political dialogue; we actually embrace it.”

Closing on an optimistic note, Griffin urged the students in the audience to go after success, declaring, “The world is always awash in challenge and its shortcomings, but no matter what anybody says, you live at the greatest moment in the history of the planet. Make the most of it.”

According to MIT’s charter, established in 1861, part of the Institute’s mission is to advance the “development and practical application of science in connection with arts, agriculture, manufactures, and commerce.” Today, the Abdul Latif Jameel Water and Food Systems Lab (J-WAFS) is one of the driving forces behind water and food-related research on campus, much of which relates to agriculture. In 2022, J-WAFS established the Water and Food Grand Challenge Grant to inspire MIT researchers to work toward a water-secure and food-secure future for our changing planet. Not unlike MIT’s Climate Grand Challenges, the J-WAFS Grand Challenge seeks to leverage multiple areas of expertise, programs, and Institute resources. The initial call for statements of interests returned 23 letters from MIT researchers spanning 18 departments, labs, and centers. J-WAFS hosted workshops for the proposers to present and discuss their initial ideas. These were winnowed down to a smaller set of invited concept papers, followed by the final proposal stage. 

Today, J-WAFS is delighted to report that the inaugural J-WAFS Grand Challenge Grant has been awarded to a team of researchers led by Professor Matt Shoulders and research scientist Robert Wilson of the Department of Chemistry. A panel of expert, external reviewers highly endorsed their proposal, which tackles a longstanding problem in crop biology — how to make photosynthesis more efficient. The team will receive $1.5 million over three years to facilitate a multistage research project that combines cutting-edge innovations in synthetic and computational biology. If successful, this project could create major benefits for agriculture and food systems worldwide.

“Food systems are a major source of global greenhouse gas emissions, and they are also increasingly vulnerable to the impacts of climate change. That’s why when we talk about climate change, we have to talk about food systems, and vice versa,” says Maria T. Zuber, MIT’s vice president for research. “J-WAFS is central to MIT’s efforts to address the interlocking challenges of climate, water, and food. This new grant program aims to catalyze innovative projects that will have real and meaningful impacts on water and food. I congratulate Professor Shoulders and the rest of the research team on being the inaugural recipients of this grant.”

Shoulders will work with Bryan Bryson, associate professor of biological engineering, as well as Bin Zhang, associate professor of chemistry, and Mary Gehring, a professor in the Department of Biology and the Whitehead Institute for Biomedical Research. Robert Wilson from the Shoulders lab will be coordinating the research effort. The team at MIT will work with outside collaborators Spencer Whitney, a professor from the Australian National University, and Ahmed Badran, an assistant professor at the Scripps Research Institute. A milestone-based collaboration will also take place with Stephen Long, a professor from the University of Illinois at Urbana-Champaign. The group consists of experts in continuous directed evolution, machine learning, molecular dynamics simulations, translational plant biochemistry, and field trials.

“This project seeks to fundamentally improve the RuBisCO enzyme that plants use to convert carbon dioxide into the energy-rich molecules that constitute our food,” says J-WAFS Director John H. Lienhard V. “This difficult problem is a true grand challenge, calling for extensive resources. With J-WAFS’ support, this long-sought goal may finally be achieved through MIT’s leading-edge research,” he adds.

RuBisCO: No, it’s not a new breakfast cereal; it just might be the key to an agricultural revolution

A growing global population, the effects of climate change, and social and political conflicts like the war in Ukraine are all threatening food supplies, particularly grain crops. Current projections estimate that crop production must increase by at least 50 percent over the next 30 years to meet food demands. One key barrier to increased crop yields is a photosynthetic enzyme called Ribulose-1,5-Bisphosphate Carboxylase/Oxygenase (RuBisCO). During photosynthesis, crops use energy gathered from light to draw carbon dioxide (CO2) from the atmosphere and transform it into sugars and cellulose for growth, a process known as carbon fixation. RuBisCO is essential for capturing the CO2 from the air to initiate conversion of CO2 into energy-rich molecules like glucose. This reaction occurs during the second stage of photosynthesis, also known as the Calvin cycle. Without RuBisCO, the chemical reactions that account for virtually all carbon acquisition in life could not occur.

Unfortunately, RuBisCO has biochemical shortcomings. Notably, the enzyme acts slowly. Many other enzymes can process a thousand molecules per second, but RuBisCO in chloroplasts fixes less than six carbon dioxide molecules per second, often limiting the rate of plant photosynthesis. Another problem is that oxygen (O2) molecules and carbon dioxide molecules are relatively similar in shape and chemical properties, and RuBisCO is unable to fully discriminate between the two. The inadvertent fixation of oxygen by RuBisCO leads to energy and carbon loss. What’s more, at higher temperatures RuBisCO reacts even more frequently with oxygen, which will contribute to decreased photosynthetic efficiency in many staple crops as our climate warms.

The scientific consensus is that genetic engineering and synthetic biology approaches could revolutionize photosynthesis and offer protection against crop losses. To date, crop RuBisCO engineering has been impaired by technological obstacles that have limited any success in significantly enhancing crop production. Excitingly, genetic engineering and synthetic biology tools are now at a point where they can be applied and tested with the aim of creating crops with new or improved biological pathways for producing more food for the growing population.

An epic plan for fighting food insecurity

The 2023 J-WAFS Grand Challenge project will use state-of-the-art, transformative protein engineering techniques drawn from biomedicine to improve the biochemistry of photosynthesis, specifically focusing on RuBisCO. Shoulders and his team are planning to build what they call the Enhanced Photosynthesis in Crops (EPiC) platform. The project will evolve and design better crop RuBisCO in the laboratory, followed by validation of the improved enzymes in plants, ultimately resulting in the deployment of enhanced RuBisCO in field trials to evaluate the impact on crop yield. 

Several recent developments make high-throughput engineering of crop RuBisCO possible. RuBisCO requires a complex chaperone network for proper assembly and function in plants. Chaperones are like helpers that guide proteins during their maturation process, shielding them from aggregation while coordinating their correct assembly. Wilson and his collaborators previously unlocked the ability to recombinantly produce plant RuBisCO outside of plant chloroplasts by reconstructing this chaperone network in Escherichia coli (E. coli). Whitney has now established that the RuBisCO enzymes from a range of agriculturally relevant crops, including potato, carrot, strawberry, and tobacco, can also be expressed using this technology. Whitney and Wilson have further developed a range of RuBisCO-dependent E. coli screens that can identify improved RuBisCO from complex gene libraries. Moreover, Shoulders and his lab have developed sophisticated in vivo mutagenesis technologies that enable efficient continuous directed evolution campaigns. Continuous directed evolution refers to a protein engineering process that can accelerate the steps of natural evolution simultaneously in an uninterrupted cycle in the lab, allowing for rapid testing of protein sequences. While Shoulders and Badran both have prior experience with cutting-edge directed evolution platforms, this will be the first time directed evolution is applied to RuBisCO from plants.

Artificial intelligence is changing the way enzyme engineering is undertaken by researchers. Principal investigators Zhang and Bryson will leverage modern computational methods to simulate the dynamics of RuBisCO structure and explore its evolutionary landscape. Specifically, Zhang will use molecular dynamics simulations to simulate and monitor the conformational dynamics of the atoms in a protein and its programmed environment over time. This approach will help the team evaluate the effect of mutations and new chemical functionalities on the properties of RuBisCO. Bryson will employ artificial intelligence and machine learning to search the RuBisCO activity landscape for optimal sequences. The computational and biological arms of the EPiC platform will work together to both validate and inform each other’s approaches to accelerate the overall engineering effort.

Shoulders and the group will deploy their designed enzymes in tobacco plants to evaluate their effects on growth and yield relative to natural RuBisCO. Gehring, a plant biologist, will assist with screening improved RuBisCO variants using the tobacco variety Nicotiana benthamianaI, where transient expression can be deployed. Transient expression is a speedy approach to test whether novel engineered RuBisCO variants can be correctly synthesized in leaf chloroplasts. Variants that pass this quality-control checkpoint at MIT will be passed to the Whitney Lab at the Australian National University for stable transformation into Nicotiana tabacum (tobacco), enabling robust measurements of photosynthetic improvement. In a final step, Professor Long at the University of Illinois at Urbana-Champaign will perform field trials of the most promising variants.

Even small improvements could have a big impact

A common criticism of efforts to improve RuBisCO is that natural evolution has not already identified a better enzyme, possibly implying that none will be found. Traditional views have speculated a catalytic trade-off between RuBisCO’s specificity factor for CO2 / O2 versus its CO2 fixation efficiency, leading to the belief that specificity factor improvements might be offset by even slower carbon fixation or vice versa. This trade-off has been suggested to explain why natural evolution has been slow to achieve a better RuBisCO. But Shoulders and the team are convinced that the EPiC platform can unlock significant overall improvements to plant RuBisCO. This view is supported by the fact that Wilson and Whitney have previously used directed evolution to improve CO2 fixation efficiency by 50 percent in RuBisCO from cyanobacteria (the ancient progenitors of plant chloroplasts) while simultaneously increasing the specificity factor. 

The EPiC researchers anticipate that their initial variants could yield 20 percent increases in RuBisCO’s specificity factor without impairing other aspects of catalysis. More sophisticated variants could lift RuBisCO out of its evolutionary trap and display attributes not currently observed in nature. “If we achieve anywhere close to such an improvement and it translates to crops, the results could help transform agriculture,” Shoulders says. “If our accomplishments are more modest, it will still recruit massive new investments to this essential field.”

Successful engineering of RuBisCO would be a scientific feat of its own and ignite renewed enthusiasm for improving plant CO2 fixation. Combined with other advances in photosynthetic engineering, such as improved light usage, a new green revolution in agriculture could be achieved. Long-term impacts of the technology’s success will be measured in improvements to crop yield and grain availability, as well as resilience against yield losses under higher field temperatures. Moreover, improved land productivity together with policy initiatives would assist in reducing the environmental footprint of agriculture. With more “crop per drop,” reductions in water consumption from agriculture would be a major boost to sustainable farming practices.

“Our collaborative team of biochemists and synthetic biologists, computational biologists, and chemists is deeply integrated with plant biologists and field trial experts, yielding a robust feedback loop for enzyme engineering,” Shoulders adds. “Together, this team will be able to make a concerted effort using the most modern, state-of-the-art techniques to engineer crop RuBisCO with an eye to helping make meaningful gains in securing a stable crop supply, hopefully with accompanying improvements in both food and water security.”

As a car travels along a narrow city street, reflections off the glossy paint or side mirrors of parked vehicles can help the driver glimpse things that would otherwise be hidden from view, like a child playing on the sidewalk behind the parked cars.

Drawing on this idea, researchers from MIT and Rice University have created a computer vision technique that leverages reflections to image the world. Their method uses reflections to turn glossy objects into “cameras,” enabling a user to see the world as if they were looking through the “lenses” of everyday objects like a ceramic coffee mug or a metallic paper weight.   

Using images of an object taken from different angles, the technique converts the surface of that object into a virtual sensor which captures reflections. The AI system maps these reflections in a way that enables it to estimate depth in the scene and capture novel views that would only be visible from the object’s perspective. One could use this technique to see around corners or beyond objects that block the observer’s view.

This method could be especially useful in autonomous vehicles. For instance, it could enable a self-driving car to use reflections from objects it passes, like lamp posts or buildings, to see around a parked truck.

“We have shown that any surface can be converted into a sensor with this formulation that converts objects into virtual pixels and virtual sensors. This can be applied in many different areas,” says Kushagra Tiwary, a graduate student in the Camera Culture Group at the Media Lab and co-lead author of a paper on this research.

Tiwary is joined on the paper by co-lead author Akshat Dave, a graduate student at Rice University; Nikhil Behari, an MIT research support associate; Tzofi Klinghoffer, an MIT graduate student; Ashok Veeraraghavan, professor of electrical and computer engineering at Rice University; and senior author Ramesh Raskar, associate professor of media arts and sciences and leader of the Camera Culture Group at MIT. The research will be presented at the Conference on Computer Vision and Pattern Recognition.

Reflecting on reflections

The heroes in crime television shows often “zoom and enhance” surveillance footage to capture reflections — perhaps those caught in a suspect’s sunglasses — that help them solve a crime. 

“In real life, exploiting these reflections is not as easy as just pushing an enhance button. Getting useful information out of these reflections is pretty hard because reflections give us a distorted view of the world,” says Dave.

This distortion depends on the shape of the object and the world that object is reflecting, both of which researchers may have incomplete information about. In addition, the glossy object may have its own color and texture that mixes with reflections. Plus, reflections are two-dimensional projections of a three-dimensional world, which makes it hard to judge depth in reflected scenes.

The researchers found a way to overcome these challenges. Their technique, known as ORCa (which stands for Objects as Radiance-Field Cameras), works in three steps. First, they take pictures of an object from many vantage points, capturing multiple reflections on the glossy object.

Then, for each image from the real camera, ORCa uses machine learning to convert the surface of the object into a virtual sensor that captures light and reflections that strike each virtual pixel on the object’s surface. Finally, the system uses virtual pixels on the object’s surface to model the 3D environment from the point of view of the object.

Catching rays

Imaging the object from many angles enables ORCa to capture multiview reflections, which the system uses to estimate depth between the glossy object and other objects in the scene, in addition to estimating the shape of the glossy object. ORCa models the scene as a 5D radiance field, which captures additional information about the intensity and direction of light rays that emanate from and strike each point in the scene.

The additional information contained in this 5D radiance field also helps ORCa accurately estimate depth. And because the scene is represented as a 5D radiance field, rather than a 2D image, the user can see hidden features that would otherwise be blocked by corners or obstructions.

In fact, once ORCa has captured this 5D radiance field, the user can put a virtual camera anywhere in the scene and synthesize what that camera would see, Dave explains. The user could also insert virtual objects into the environment or change the appearance of an object, such as from ceramic to metallic.

“It was especially challenging to go from a 2D image to a 5D environment. You have to make sure that mapping works and is physically accurate, so it is based on how light travels in space and how light interacts with the environment. We spent a lot of time thinking about how we can model a surface,” Tiwary says.

Accurate estimations

The researchers evaluated their technique by comparing it with other methods that model reflections, which is a slightly different task than ORCa performs. Their method performed well at separating out the true color of an object from the reflections, and it outperformed the baselines by extracting more accurate object geometry and textures.

They compared the system’s depth estimations with simulated ground truth data on the actual distance between objects in the scene and found ORCa’s predictions to be reliable.   

“Consistently, with ORCa, it not only estimates the environment accurately as a 5D image, but to achieve that, in the intermediate steps, it also does a good job estimating the shape of the object and separating the reflections from the object texture,” Dave says.

Building off of this proof-of-concept, the researchers want to apply this technique to drone imaging. ORCa could use faint reflections from objects a drone flies over to reconstruct a scene from the ground. They also want to enhance ORCa so it can utilize other cues, such as shadows, to reconstruct hidden information, or combine reflections from two objects to image new parts of a scene.

“Estimating specular reflections is really important for seeing around corners, and this is the next natural step to see around corners using faint reflections in the scene,” says Raskar.

“Ordinarily, shiny objects are difficult for vision systems to handle. This paper is very creative because it turns the longstanding weakness of object shininess into an advantage. By exploiting environment reflections off a shiny object, the paper is not only able to see hidden parts of the scene, but also understand how the scene is lit. This enables applications in 3D perception that include, but are not limited to, an ability to composite virtual objects into real scenes in ways that appear seamless, even in challenging lighting conditions,” says Achuta Kadambi, assistant professor of electrical engineering and computer science at the University of California at Los Angeles, who was not involved with this work. “One reason that others have not been able to use shiny objects in this fashion is that most prior works require surfaces with known geometry or texture. The authors have derived an intriguing, new formulation that does not require such knowledge.”

The research was supported, in part, by the Intelligence Advanced Research Projects Activity and the National Science Foundation.

Imagine sitting on a park bench, watching someone stroll by. While the scene may constantly change as the person walks, the human brain can transform that dynamic visual information into a more stable representation over time. This ability, known as perceptual straightening, helps us predict the walking person’s trajectory.

Unlike humans, computer vision models don’t typically exhibit perceptual straightness, so they learn to represent visual information in a highly unpredictable way. But if machine-learning models had this ability, it might enable them to better estimate how objects or people will move.

MIT researchers have discovered that a specific training method can help computer vision models learn more perceptually straight representations, like humans do. Training involves showing a machine-learning model millions of examples so it can learn a task.

The researchers found that training computer vision models using a technique called adversarial training, which makes them less reactive to tiny errors added to images, improves the models’ perceptual straightness.

The team also discovered that perceptual straightness is affected by the task one trains a model to perform. Models trained to perform abstract tasks, like classifying images, learn more perceptually straight representations than those trained to perform more fine-grained tasks, like assigning every pixel in an image to a category.   

For example, the nodes within the model have internal activations that represent “dog,” which allow the model to detect a dog when it sees any image of a dog. Perceptually straight representations retain a more stable “dog” representation when there are small changes in the image. This makes them more robust.

By gaining a better understanding of perceptual straightness in computer vision, the researchers hope to uncover insights that could help them develop models that make more accurate predictions. For instance, this property might improve the safety of autonomous vehicles that use computer vision models to predict the trajectories of pedestrians, cyclists, and other vehicles.

“One of the take-home messages here is that taking inspiration from biological systems, such as human vision, can both give you insight about why certain things work the way that they do and also inspire ideas to improve neural networks,” says Vasha DuTell, an MIT postdoc and co-author of a paper exploring perceptual straightness in computer vision.

Joining DuTell on the paper are lead author Anne Harrington, a graduate student in the Department of Electrical Engineering and Computer Science (EECS); Ayush Tewari, a postdoc; Mark Hamilton, a graduate student; Simon Stent, research manager at Woven Planet; Ruth Rosenholtz, principal research scientist in the Department of Brain and Cognitive Sciences and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author William T. Freeman, the Thomas and Gerd Perkins Professor of Electrical Engineering and Computer Science and a member of CSAIL. The research is being presented at the International Conference on Learning Representations.

Studying straightening

After reading a 2019 paper from a team of New York University researchers about perceptual straightness in humans, DuTell, Harrington, and their colleagues wondered if that property might be useful in computer vision models, too.

They set out to determine whether different types of computer vision models straighten the visual representations they learn. They fed each model frames of a video and then examined the representation at different stages in its learning process.

If the model’s representation changes in a predictable way across the frames of the video, that model is straightening. At the end, its output representation should be more stable than the input representation.

“You can think of the representation as a line, which starts off really curvy. A model that straightens can take that curvy line from the video and straighten it out through its processing steps,” DuTell explains.

Most models they tested didn’t straighten. Of the few that did, those which straightened most effectively had been trained for classification tasks using the technique known as adversarial training.

Adversarial training involves subtly modifying images by slightly changing each pixel. While a human wouldn’t notice the difference, these minor changes can fool a machine so it misclassifies the image. Adversarial training makes the model more robust, so it won’t be tricked by these manipulations.

Because adversarial training teaches the model to be less reactive to slight changes in images, this helps it learn a representation that is more predictable over time, Harrington explains.

“People have already had this idea that adversarial training might help you get your model to be more like a human, and it was interesting to see that carry over to another property that people hadn’t tested before,” she says.

But the researchers found that adversarially trained models only learn to straighten when they are trained for broad tasks, like classifying entire images into categories. Models tasked with segmentation — labeling every pixel in an image as a certain class — did not straighten, even when they were adversarially trained.

Consistent classification

The researchers tested these image classification models by showing them videos. They found that the models which learned more perceptually straight representations tended to correctly classify objects in the videos more consistently.

“To me, it is amazing that these adversarially trained models, which have never even seen a video and have never been trained on temporal data, still show some amount of straightening,” DuTell says.

The researchers don’t know exactly what about the adversarial training process enables a computer vision model to straighten, but their results suggest that stronger training schemes cause the models to straighten more, she explains.

Building off this work, the researchers want to use what they learned to create new training schemes that would explicitly give a model this property. They also want to dig deeper into adversarial training to understand why this process helps a model straighten.

“From a biological standpoint, adversarial training doesn’t necessarily make sense. It’s not how humans understand the world. There are still a lot of questions about why this training process seems to help models act more like humans,” Harrington says.

“Understanding the representations learned by deep neural networks is critical to improve properties such as robustness and generalization,” says Bill Lotter, assistant professor at the Dana-Farber Cancer Institute and Harvard Medical School, who was not involved with this research. “Harrington et al. perform an extensive evaluation of how the representations of computer vision models change over time when processing natural videos, showing that the curvature of these trajectories varies widely depending on model architecture, training properties, and task. These findings can inform the development of improved models and also offer insights into biological visual processing.”

“The paper confirms that straightening natural videos is a fairly unique property displayed by the human visual system. Only adversarially trained networks display it, which provides an interesting connection with another signature of human perception: its robustness to various image transformations, whether natural or artificial,” says Olivier Hénaff, a research scientist at DeepMind, who was not involved with this research. “That even adversarially trained scene segmentation models do not straighten their inputs raises important questions for future work: Do humans parse natural scenes in the same way as computer vision models? How to represent and predict the trajectories of objects in motion while remaining sensitive to their spatial detail? In connecting the straightening hypothesis with other aspects of visual behavior, the paper lays the groundwork for more unified theories of perception.”

The research is funded, in part, by the Toyota Research Institute, the MIT CSAIL METEOR Fellowship, the National Science Foundation, the U.S. Air Force Research Laboratory, and the U.S. Air Force Artificial Intelligence Accelerator.

Researchers often use simulations when designing new algorithms, since testing ideas in the real world can be both costly and risky. But since it’s impossible to capture every detail of a complex system in a simulation, they typically collect a small amount of real data that they replay while simulating the components they want to study.

Known as trace-driven simulation (the small pieces of real data are called traces), this method sometimes results in biased outcomes. This means researchers might unknowingly choose an algorithm that is not the best one they evaluated, and which will perform worse on real data than the simulation predicted that it should.

MIT researchers have developed a new method that eliminates this source of bias in trace-driven simulation. By enabling unbiased trace-driven simulations, the new technique could help researchers design better algorithms for a variety of applications, including improving video quality on the internet and increasing the performance of data processing systems.

The researchers’ machine-learning algorithm draws on the principles of causality to learn how the data traces were affected by the behavior of the system. In this way, they can replay the correct, unbiased version of the trace during the simulation.

When compared to a previously developed trace-driven simulator, the researchers’ simulation method correctly predicted which newly designed algorithm would be best for video streaming — meaning the one that led to less rebuffering and higher visual quality. Existing simulators that do not account for bias would have pointed researchers to a worse-performing algorithm.

“Data are not the only thing that matter. The story behind how the data are generated and collected is also important. If you want to answer a counterfactual question, you need to know the underlying data generation story so you only intervene on those things that you really want to simulate,” says Arash Nasr-Esfahany, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper on this new technique.

He is joined on the paper by co-lead authors and fellow EECS graduate students Abdullah Alomar and Pouya Hamadanian; recent graduate student Anish Agarwal PhD ’21; and senior authors Mohammad Alizadeh, an associate professor of electrical engineering and computer science; and Devavrat Shah, the Andrew and Erna Viterbi Professor in EECS and a member of the Institute for Data, Systems, and Society and of the Laboratory for Information and Decision Systems. The research was recently presented at the USENIX Symposium on Networked Systems Design and Implementation.

Specious simulations

The MIT researchers studied trace-driven simulation in the context of video streaming applications.

In video streaming, an adaptive bitrate algorithm continually decides the video quality, or bitrate, to transfer to a device based on real-time data on the user’s bandwidth. To test how different adaptive bitrate algorithms impact network performance, researchers can collect real data from users during a video stream for a trace-driven simulation.

They use these traces to simulate what would have happened to network performance had the platform used a different adaptive bitrate algorithm in the same underlying conditions.

Researchers have traditionally assumed that trace data are exogenous, meaning they aren’t affected by factors that are changed during the simulation. They would assume that, during the period when they collected the network performance data, the choices the bitrate adaptation algorithm made did not affect those data.

But this is often a false assumption that results in biases about the behavior of new algorithms, making the simulation invalid, Alizadeh explains.

“We recognized, and others have recognized, that this way of doing simulation can induce errors. But I don’t think people necessarily knew how significant those errors could be,” he says.

To develop a solution, Alizadeh and his collaborators framed the issue as a causal inference problem. To collect an unbiased trace, one must understand the different causes that affect the observed data. Some causes are intrinsic to a system, while others are affected by the actions being taken.

In the video streaming example, network performance is affected by the choices the bitrate adaptation algorithm made — but it’s also affected by intrinsic elements, like network capacity.

“Our task is to disentangle these two effects, to try to understand what aspects of the behavior we are seeing are intrinsic to the system and how much of what we are observing is based on the actions that were taken. If we can disentangle these two effects, then we can do unbiased simulations,” he says.

Learning from data

But researchers often cannot directly observe intrinsic properties. This is where the new tool, called CausalSim, comes in. The algorithm can learn the underlying characteristics of a system using only the trace data.

CausalSim takes trace data that were collected through a randomized control trial, and estimates the underlying functions that produced those data. The model tells the researchers, under the exact same underlying conditions that a user experienced, how a new algorithm would change the outcome.

Using a typical trace-driven simulator, bias might lead a researcher to select a worse-performing algorithm, even though the simulation indicates it should be better. CausalSim helps researchers select the best algorithm that was tested.

The MIT researchers observed this in practice. When they used CausalSim to design an improved bitrate adaptation algorithm, it led them to select a new variant that had a stall rate that was nearly 1.4 times lower than a well-accepted competing algorithm, while achieving the same video quality. The stall rate is the amount of time a user spent rebuffering the video.

By contrast, an expert-designed trace-driven simulator predicted the opposite. It indicated that this new variant should cause a stall rate that was nearly 1.3 times higher. The researchers tested the algorithm on real-world video streaming and confirmed that CausalSim was correct.

“The gains we were getting in the new variant were very close to CausalSim’s prediction, while the expert simulator was way off. This is really exciting because this expert-designed simulator has been used in research for the past decade. If CausalSim can so clearly be better than this, who knows what we can do with it?” says Hamadanian.

During a 10-month experiment, CausalSim consistently improved simulation accuracy, resulting in algorithms that made about half as many errors as those designed using baseline methods.

In the future, the researchers want to apply CausalSim to situations where randomized control trial data are not available or where it is especially difficult to recover the causal dynamics of the system. They also want to explore how to design and monitor systems to make them more amenable to causal analysis.

When medical companies manufacture the pills and tablets that treat any number of illnesses, aches, and pains, they need to isolate the active pharmaceutical ingredient from a suspension and dry it. The process requires a human operator to monitor an industrial dryer, agitate the material, and watch for the compound to take on the right qualities for compressing into medicine. The job depends heavily on the operator’s observations.   

Methods for making that process less subjective and a lot more efficient are the subject of a recent Nature Communications paper authored by researchers at MIT and Takeda. The paper’s authors devise a way to use physics and machine learning to categorize the rough surfaces that characterize particles in a mixture. The technique, which uses a physics-enhanced autocorrelation-based estimator (PEACE), could change pharmaceutical manufacturing processes for pills and powders, increasing efficiency and accuracy and resulting in fewer failed batches of pharmaceutical products.  

“Failed batches or failed steps in the pharmaceutical process are very serious,” says Allan Myerson, a professor of practice in the MIT Department of Chemical Engineering and one of the study’s authors. “Anything that improves the reliability of the pharmaceutical manufacturing, reduces time, and improves compliance is a big deal.”

The team’s work is part of an ongoing collaboration between Takeda and MIT, launched in 2020. The MIT-Takeda Program aims to leverage the experience of both MIT and Takeda to solve problems at the intersection of medicine, artificial intelligence, and health care.

In pharmaceutical manufacturing, determining whether a compound is adequately mixed and dried ordinarily requires stopping an industrial-sized dryer and taking samples off the manufacturing line for testing. Researchers at Takeda thought artificial intelligence could improve the task and reduce stoppages that slow down production. Originally the research team planned to use videos to train a computer model to replace a human operator. But determining which videos to use to train the model still proved too subjective. Instead, the MIT-Takeda team decided to illuminate particles with a laser during filtration and drying, and measure particle size distribution using physics and machine learning. 

“We just shine a laser beam on top of this drying surface and observe,” says Qihang Zhang, a doctoral student in MIT’s Department of Electrical Engineering and Computer Science and the study’s first author. 

A physics-derived equation describes the interaction between the laser and the mixture, while machine learning characterizes the particle sizes. The process doesn’t require stopping and starting the process, which means the entire job is more secure and more efficient than standard operating procedure, according to George Barbastathis, professor of mechanical engineering at MIT and corresponding author of the study.

The machine learning algorithm also does not require many datasets to learn its job, because the physics allows for speedy training of the neural network.

“We utilize the physics to compensate for the lack of training data, so that we can train the neural network in an efficient way,” says Zhang. “Only a tiny amount of experimental data is enough to get a good result.”

Today, the only inline processes used for particle measurements in the pharmaceutical industry are for slurry products, where crystals float in a liquid. There is no method for measuring particles within a powder during mixing. Powders can be made from slurries, but when a liquid is filtered and dried its composition changes, requiring new measurements. In addition to making the process quicker and more efficient, using the PEACE mechanism makes the job safer because it requires less handling of potentially highly potent materials, the authors say. 

The ramifications for pharmaceutical manufacturing could be significant, allowing drug production to be more efficient, sustainable, and cost-effective, by reducing the number of experiments companies need to conduct when making products. Monitoring the characteristics of a drying mixture is an issue the industry has long struggled with, according to Charles Papageorgiou, the director of Takeda’s Process Chemistry Development group and one of the study’s authors. 

“It is a problem that a lot of people are trying to solve, and there isn’t a good sensor out there,” says Papageorgiou. “This is a pretty big step change, I think, with respect to being able to monitor, in real time, particle size distribution.”

Papageorgiou said that the mechanism could have applications in other industrial pharmaceutical operations. At some point, the laser technology may be able to train video imaging, allowing manufacturers to use a camera for analysis rather than laser measurements. The company is now working to assess the tool on different compounds in its lab. 

The results come directly from collaboration between Takeda and three MIT departments: Mechanical Engineering, Chemical Engineering, and Electrical Engineering and Computer Science. Over the last three years, researchers at MIT and Takeda have worked together on 19 projects focused on applying machine learning and artificial intelligence to problems in the health-care and medical industry as part of the MIT-Takeda Program. 

Often, it can take years for academic research to translate to industrial processes. But researchers are hopeful that direct collaboration could shorten that timeline. Takeda is a walking distance away from MIT’s campus, which allowed researchers to set up tests in the company’s lab, and real-time feedback from Takeda helped MIT researchers structure their research based on the company’s equipment and operations. 

Combining the expertise and mission of both entities helps researchers ensure their experimental results will have real-world implications. The team has already filed for two patents and has plans to file for a third.  

Martin Wainwright, the Cecil H. Green Professor in MIT’s departments of Electrical Engineering and Computer Science (EECS) and Mathematics, has been named the new director of the Institute for Data, Systems, and Society (IDSS), effective July 1.

“Martin is a widely recognized leader in statistics and machine learning — both in research and in education. In taking on this leadership role in the college, Martin will work to build up the human and institutional behavior component of IDSS, while strengthening initiatives in both policy and statistics, and collaborations within the institute, across MIT, and beyond,” says Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing and the Henry Ellis Warren Professor of Electrical Engineering and Computer Science. “I look forward to working with him and supporting his efforts in this next chapter for IDSS.”

“Martin holds a strong belief in the value of theoretical, experimental, and computational approaches to research and in facilitating connections between them. He also places much importance in having practical, as well as academic, impact,” says Asu Ozdaglar, deputy dean of academics for the MIT Schwarzman College of Computing, department head of EECS, and the MathWorks Professor of Electrical Engineering and Computer Science. “As the new director of IDSS, he will undoubtedly bring these tenets to the role in advancing the mission of IDSS and helping to shape its future.”

A principal investigator in the Laboratory for Information and Decision Systems and the Statistics and Data Science Center, Wainwright joined the MIT faculty in July 2022 from the University of California at Berkeley, where he held the Howard Friesen Chair with a joint appointment between the departments of Electrical Engineering and Computer Science and Statistics.

Wainwright received his bachelor’s degree in mathematics from the University of Waterloo, Canada, and doctoral degree in electrical engineering and computer science from MIT. He has received a number of awards and recognition, including an Alfred P. Sloan Foundation Fellowship, and best paper awards from the IEEE Signal Processing Society, IEEE Communications Society, and IEEE Information Theory and Communication Societies. He has also been honored with the Medallion Lectureship and Award from the Institute of Mathematical Statistics, and the COPSS Presidents’ Award from the Joint Statistical Societies. He was a section lecturer with the International Congress of Mathematicians in 2014 and received the Blackwell Award from the Institute of Mathematical Statistics in 2017.

He is the author of “High-dimensional Statistics: A Non-Asymptotic Viewpoint” (Cambridge University Press, 2019), and is coauthor on several books, including on graphical models and on sparse statistical modeling.

Wainwright succeeds Munther Dahleh, the William A. Coolidge Professor in EECS, who has helmed IDSS since its founding in 2015.

“I am grateful to Munther and thank him for his leadership of IDSS. As the founding director, he has led the creation of a remarkable new part of MIT,” says Huttenlocher.