Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 125 Shares129 Views

      in Data Management & Statistics

      How symmetry can come to the aid of machine learning

      by

      Behrooz Tahmasebi — an MIT PhD student in the Department of Electrical Engineering and Computer Science (EECS) and an affiliate of the Computer Science and Artificial Intelligence Laboratory (CSAIL) — was taking a mathematics course on differential equations in late 2021 when a glimmer of inspiration struck. In that class, he learned for the first time about Weyl’s law, which had been formulated 110 years earlier by the German mathematician Hermann Weyl. Tahmasebi realized it might have some relevance to the computer science problem he was then wrestling with, even though the connection appeared — on the surface — to be thin, at best. Weyl’s law, he says, provides a formula that measures the complexity of the spectral information, or data, contained within the fundamental frequencies of a drum head or guitar string.

      Tahmasebi was, at the same time, thinking about measuring the complexity of the input data to a neural network, wondering whether that complexity could be reduced by taking into account some of the symmetries inherent to the dataset. Such a reduction, in turn, could facilitate — as well as speed up — machine learning processes.

      Weyl’s law, conceived about a century before the boom in machine learning, had traditionally been applied to very different physical situations — such as those concerning the vibrations of a string or the spectrum of electromagnetic (black-body) radiation given off by a heated object. Nevertheless, Tahmasebi believed that a customized version of that law might help with the machine learning problem he was pursuing. And if the approach panned out, the payoff could be considerable.

      He spoke with his advisor, Stefanie Jegelka — an associate professor in EECS and affiliate of CSAIL and the MIT Institute for Data, Systems, and Society — who believed the idea was definitely worth looking into. As Tahmasebi saw it, Weyl’s law had to do with gauging the complexity of data, and so did this project. But Weyl’s law, in its original form, said nothing about symmetry.

      He and Jegelka have now succeeded in modifying Weyl’s law so that symmetry can be factored into the assessment of a dataset’s complexity. “To the best of my knowledge,” Tahmasebi says, “this is the first time Weyl’s law has been used to determine how machine learning can be enhanced by symmetry.”

      The paper he and Jegelka wrote earned a “Spotlight” designation when it was presented at the December 2023 conference on Neural Information Processing Systems — widely regarded as the world’s top conference on machine learning.

      This work, comments Soledad Villar, an applied mathematician at Johns Hopkins University, “shows that models that satisfy the symmetries of the problem are not only correct but also can produce predictions with smaller errors, using a small amount of training points. [This] is especially important in scientific domains, like computational chemistry, where training data can be scarce.”

      In their paper, Tahmasebi and Jegelka explored the ways in which symmetries, or so-called “invariances,” could benefit machine learning. Suppose, for example, the goal of a particular computer run is to pick out every image that contains the numeral 3. That task can be a lot easier, and go a lot quicker, if the algorithm can identify the 3 regardless of where it is placed in the box — whether it’s exactly in the center or off to the side — and whether it is pointed right-side up, upside down, or oriented at a random angle. An algorithm equipped with the latter capability can take advantage of the symmetries of translation and rotations, meaning that a 3, or any other object, is not changed in itself by altering its position or by rotating it around an arbitrary axis. It is said to be invariant to those shifts. The same logic can be applied to algorithms charged with identifying dogs or cats. A dog is a dog is a dog, one might say, irrespective of how it is embedded within an image. 

      The point of the entire exercise, the authors explain, is to exploit a dataset’s intrinsic symmetries in order to reduce the complexity of machine learning tasks. That, in turn, can lead to a reduction in the amount of data needed for learning. Concretely, the new work answers the question: How many fewer data are needed to train a machine learning model if the data contain symmetries?

      There are two ways of achieving a gain, or benefit, by capitalizing on the symmetries present. The first has to do with the size of the sample to be looked at. Let’s imagine that you are charged, for instance, with analyzing an image that has mirror symmetry — the right side being an exact replica, or mirror image, of the left. In that case, you don’t have to look at every pixel; you can get all the information you need from half of the image — a factor of two improvement. If, on the other hand, the image can be partitioned into 10 identical parts, you can get a factor of 10 improvement. This kind of boosting effect is linear.

      To take another example, imagine you are sifting through a dataset, trying to find sequences of blocks that have seven different colors — black, blue, green, purple, red, white, and yellow. Your job becomes much easier if you don’t care about the order in which the blocks are arranged. If the order mattered, there would be 5,040 different combinations to look for. But if all you care about are sequences of blocks in which all seven colors appear, then you have reduced the number of things — or sequences — you are searching for from 5,040 to just one.

      Tahmasebi and Jegelka discovered that it is possible to achieve a different kind of gain — one that is exponential — that can be reaped for symmetries that operate over many dimensions. This advantage is related to the notion that the complexity of a learning task grows exponentially with the dimensionality of the data space. Making use of a multidimensional symmetry can therefore yield a disproportionately large return. “This is a new contribution that is basically telling us that symmetries of higher dimension are more important because they can give us an exponential gain,” Tahmasebi says. 

      The NeurIPS 2023 paper that he wrote with Jegelka contains two theorems that were proved mathematically. “The first theorem shows that an improvement in sample complexity is achievable with the general algorithm we provide,” Tahmasebi says. The second theorem complements the first, he added, “showing that this is the best possible gain you can get; nothing else is achievable.”

      He and Jegelka have provided a formula that predicts the gain one can obtain from a particular symmetry in a given application. A virtue of this formula is its generality, Tahmasebi notes. “It works for any symmetry and any input space.” It works not only for symmetries that are known today, but it could also be applied in the future to symmetries that are yet to be discovered. The latter prospect is not too farfetched to consider, given that the search for new symmetries has long been a major thrust in physics. That suggests that, as more symmetries are found, the methodology introduced by Tahmasebi and Jegelka should only get better over time.

      According to Haggai Maron, a computer scientist at Technion (the Israel Institute of Technology) and NVIDIA who was not involved in the work, the approach presented in the paper “diverges substantially from related previous works, adopting a geometric perspective and employing tools from differential geometry. This theoretical contribution lends mathematical support to the emerging subfield of ‘Geometric Deep Learning,’ which has applications in graph learning, 3D data, and more. The paper helps establish a theoretical basis to guide further developments in this rapidly expanding research area.” More

    • 138 Shares199 Views

      in Data Management & Statistics

      New hope for early pancreatic cancer intervention via AI-based risk prediction

      by

      The first documented case of pancreatic cancer dates back to the 18th century. Since then, researchers have undertaken a protracted and challenging odyssey to understand the elusive and deadly disease. To date, there is no better cancer treatment than early intervention. Unfortunately, the pancreas, nestled deep within the abdomen, is particularly elusive for early detection. 

      MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) scientists, alongside Limor Appelbaum, a staff scientist in the Department of Radiation Oncology at Beth Israel Deaconess Medical Center (BIDMC), were eager to better identify potential high-risk patients. They set out to develop two machine-learning models for early detection of pancreatic ductal adenocarcinoma (PDAC), the most common form of the cancer. To access a broad and diverse database, the team synced up with a federated network company, using electronic health record data from various institutions across the United States. This vast pool of data helped ensure the models’ reliability and generalizability, making them applicable across a wide range of populations, geographical locations, and demographic groups.

      The two models — the “PRISM” neural network, and the logistic regression model (a statistical technique for probability), outperformed current methods. The team’s comparison showed that while standard screening criteria identify about 10 percent of PDAC cases using a five-times higher relative risk threshold, Prism can detect 35 percent of PDAC cases at this same threshold. 

      Using AI to detect cancer risk is not a new phenomena — algorithms analyze mammograms, CT scans for lung cancer, and assist in the analysis of Pap smear tests and HPV testing, to name a few applications. “The PRISM models stand out for their development and validation on an extensive database of over 5 million patients, surpassing the scale of most prior research in the field,” says Kai Jia, an MIT PhD student in electrical engineering and computer science (EECS), MIT CSAIL affiliate, and first author on an open-access paper in eBioMedicine outlining the new work. “The model uses routine clinical and lab data to make its predictions, and the diversity of the U.S. population is a significant advancement over other PDAC models, which are usually confined to specific geographic regions, like a few health-care centers in the U.S. Additionally, using a unique regularization technique in the training process enhanced the models’ generalizability and interpretability.” 

      “This report outlines a powerful approach to use big data and artificial intelligence algorithms to refine our approach to identifying risk profiles for cancer,” says David Avigan, a Harvard Medical School professor and the cancer center director and chief of hematology and hematologic malignancies at BIDMC, who was not involved in the study. “This approach may lead to novel strategies to identify patients with high risk for malignancy that may benefit from focused screening with the potential for early intervention.” 

      Prismatic perspectives

      The journey toward the development of PRISM began over six years ago, fueled by firsthand experiences with the limitations of current diagnostic practices. “Approximately 80-85 percent of pancreatic cancer patients are diagnosed at advanced stages, where cure is no longer an option,” says senior author Appelbaum, who is also a Harvard Medical School instructor as well as radiation oncologist. “This clinical frustration sparked the idea to delve into the wealth of data available in electronic health records (EHRs).”The CSAIL group’s close collaboration with Appelbaum made it possible to understand the combined medical and machine learning aspects of the problem better, eventually leading to a much more accurate and transparent model. “The hypothesis was that these records contained hidden clues — subtle signs and symptoms that could act as early warning signals of pancreatic cancer,” she adds. “This guided our use of federated EHR networks in developing these models, for a scalable approach for deploying risk prediction tools in health care.”Both PrismNN and PrismLR models analyze EHR data, including patient demographics, diagnoses, medications, and lab results, to assess PDAC risk. PrismNN uses artificial neural networks to detect intricate patterns in data features like age, medical history, and lab results, yielding a risk score for PDAC likelihood. PrismLR uses logistic regression for a simpler analysis, generating a probability score of PDAC based on these features. Together, the models offer a thorough evaluation of different approaches in predicting PDAC risk from the same EHR data.

      One paramount point for gaining the trust of physicians, the team notes, is better understanding how the models work, known in the field as interpretability. The scientists pointed out that while logistic regression models are inherently easier to interpret, recent advancements have made deep neural networks somewhat more transparent. This helped the team to refine the thousands of potentially predictive features derived from EHR of a single patient to approximately 85 critical indicators. These indicators, which include patient age, diabetes diagnosis, and an increased frequency of visits to physicians, are automatically discovered by the model but match physicians’ understanding of risk factors associated with pancreatic cancer. 

      The path forward

      Despite the promise of the PRISM models, as with all research, some parts are still a work in progress. U.S. data alone are the current diet for the models, necessitating testing and adaptation for global use. The path forward, the team notes, includes expanding the model’s applicability to international datasets and integrating additional biomarkers for more refined risk assessment.

      “A subsequent aim for us is to facilitate the models’ implementation in routine health care settings. The vision is to have these models function seamlessly in the background of health care systems, automatically analyzing patient data and alerting physicians to high-risk cases without adding to their workload,” says Jia. “A machine-learning model integrated with the EHR system could empower physicians with early alerts for high-risk patients, potentially enabling interventions well before symptoms manifest. We are eager to deploy our techniques in the real world to help all individuals enjoy longer, healthier lives.” 

      Jia wrote the paper alongside Applebaum and MIT EECS Professor and CSAIL Principal Investigator Martin Rinard, who are both senior authors of the paper. Researchers on the paper were supported during their time at MIT CSAIL, in part, by the Defense Advanced Research Projects Agency, Boeing, the National Science Foundation, and Aarno Labs. TriNetX provided resources for the project, and the Prevent Cancer Foundation also supported the team. More

    • 113 Shares119 Views

      in Data Management & Statistics

      Image recognition accuracy: An unseen challenge confounding today’s AI

      by

      Imagine you are scrolling through the photos on your phone and you come across an image that at first you can’t recognize. It looks like maybe something fuzzy on the couch; could it be a pillow or a coat? After a couple of seconds it clicks — of course! That ball of fluff is your friend’s cat, Mocha. While some of your photos could be understood in an instant, why was this cat photo much more difficult?

      MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers were surprised to find that despite the critical importance of understanding visual data in pivotal areas ranging from health care to transportation to household devices, the notion of an image’s recognition difficulty for humans has been almost entirely ignored. One of the major drivers of progress in deep learning-based AI has been datasets, yet we know little about how data drives progress in large-scale deep learning beyond that bigger is better.

      In real-world applications that require understanding visual data, humans outperform object recognition models despite the fact that models perform well on current datasets, including those explicitly designed to challenge machines with debiased images or distribution shifts. This problem persists, in part, because we have no guidance on the absolute difficulty of an image or dataset. Without controlling for the difficulty of images used for evaluation, it’s hard to objectively assess progress toward human-level performance, to cover the range of human abilities, and to increase the challenge posed by a dataset.

      To fill in this knowledge gap, David Mayo, an MIT PhD student in electrical engineering and computer science and a CSAIL affiliate, delved into the deep world of image datasets, exploring why certain images are more difficult for humans and machines to recognize than others. “Some images inherently take longer to recognize, and it’s essential to understand the brain’s activity during this process and its relation to machine learning models. Perhaps there are complex neural circuits or unique mechanisms missing in our current models, visible only when tested with challenging visual stimuli. This exploration is crucial for comprehending and enhancing machine vision models,” says Mayo, a lead author of a new paper on the work.

      This led to the development of a new metric, the “minimum viewing time” (MVT), which quantifies the difficulty of recognizing an image based on how long a person needs to view it before making a correct identification. Using a subset of ImageNet, a popular dataset in machine learning, and ObjectNet, a dataset designed to test object recognition robustness, the team showed images to participants for varying durations from as short as 17 milliseconds to as long as 10 seconds, and asked them to choose the correct object from a set of 50 options. After over 200,000 image presentation trials, the team found that existing test sets, including ObjectNet, appeared skewed toward easier, shorter MVT images, with the vast majority of benchmark performance derived from images that are easy for humans.

      The project identified interesting trends in model performance — particularly in relation to scaling. Larger models showed considerable improvement on simpler images but made less progress on more challenging images. The CLIP models, which incorporate both language and vision, stood out as they moved in the direction of more human-like recognition.

      “Traditionally, object recognition datasets have been skewed towards less-complex images, a practice that has led to an inflation in model performance metrics, not truly reflective of a model’s robustness or its ability to tackle complex visual tasks. Our research reveals that harder images pose a more acute challenge, causing a distribution shift that is often not accounted for in standard evaluations,” says Mayo. “We released image sets tagged by difficulty along with tools to automatically compute MVT, enabling MVT to be added to existing benchmarks and extended to various applications. These include measuring test set difficulty before deploying real-world systems, discovering neural correlates of image difficulty, and advancing object recognition techniques to close the gap between benchmark and real-world performance.”

      “One of my biggest takeaways is that we now have another dimension to evaluate models on. We want models that are able to recognize any image even if — perhaps especially if — it’s hard for a human to recognize. We’re the first to quantify what this would mean. Our results show that not only is this not the case with today’s state of the art, but also that our current evaluation methods don’t have the ability to tell us when it is the case because standard datasets are so skewed toward easy images,” says Jesse Cummings, an MIT graduate student in electrical engineering and computer science and co-first author with Mayo on the paper.

      From ObjectNet to MVT

      A few years ago, the team behind this project identified a significant challenge in the field of machine learning: Models were struggling with out-of-distribution images, or images that were not well-represented in the training data. Enter ObjectNet, a dataset comprised of images collected from real-life settings. The dataset helped illuminate the performance gap between machine learning models and human recognition abilities, by eliminating spurious correlations present in other benchmarks — for example, between an object and its background. ObjectNet illuminated the gap between the performance of machine vision models on datasets and in real-world applications, encouraging use for many researchers and developers — which subsequently improved model performance.

      Fast forward to the present, and the team has taken their research a step further with MVT. Unlike traditional methods that focus on absolute performance, this new approach assesses how models perform by contrasting their responses to the easiest and hardest images. The study further explored how image difficulty could be explained and tested for similarity to human visual processing. Using metrics like c-score, prediction depth, and adversarial robustness, the team found that harder images are processed differently by networks. “While there are observable trends, such as easier images being more prototypical, a comprehensive semantic explanation of image difficulty continues to elude the scientific community,” says Mayo.

      In the realm of health care, for example, the pertinence of understanding visual complexity becomes even more pronounced. The ability of AI models to interpret medical images, such as X-rays, is subject to the diversity and difficulty distribution of the images. The researchers advocate for a meticulous analysis of difficulty distribution tailored for professionals, ensuring AI systems are evaluated based on expert standards, rather than layperson interpretations.

      Mayo and Cummings are currently looking at neurological underpinnings of visual recognition as well, probing into whether the brain exhibits differential activity when processing easy versus challenging images. The study aims to unravel whether complex images recruit additional brain areas not typically associated with visual processing, hopefully helping demystify how our brains accurately and efficiently decode the visual world.

      Toward human-level performance

      Looking ahead, the researchers are not only focused on exploring ways to enhance AI’s predictive capabilities regarding image difficulty. The team is working on identifying correlations with viewing-time difficulty in order to generate harder or easier versions of images.

      Despite the study’s significant strides, the researchers acknowledge limitations, particularly in terms of the separation of object recognition from visual search tasks. The current methodology does concentrate on recognizing objects, leaving out the complexities introduced by cluttered images.

      “This comprehensive approach addresses the long-standing challenge of objectively assessing progress towards human-level performance in object recognition and opens new avenues for understanding and advancing the field,” says Mayo. “With the potential to adapt the Minimum Viewing Time difficulty metric for a variety of visual tasks, this work paves the way for more robust, human-like performance in object recognition, ensuring that models are truly put to the test and are ready for the complexities of real-world visual understanding.”

      “This is a fascinating study of how human perception can be used to identify weaknesses in the ways AI vision models are typically benchmarked, which overestimate AI performance by concentrating on easy images,” says Alan L. Yuille, Bloomberg Distinguished Professor of Cognitive Science and Computer Science at Johns Hopkins University, who was not involved in the paper. “This will help develop more realistic benchmarks leading not only to improvements to AI but also make fairer comparisons between AI and human perception.” 

      “It’s widely claimed that computer vision systems now outperform humans, and on some benchmark datasets, that’s true,” says Anthropic technical staff member Simon Kornblith PhD ’17, who was also not involved in this work. “However, a lot of the difficulty in those benchmarks comes from the obscurity of what’s in the images; the average person just doesn’t know enough to classify different breeds of dogs. This work instead focuses on images that people can only get right if given enough time. These images are generally much harder for computer vision systems, but the best systems are only a bit worse than humans.”

      Mayo, Cummings, and Xinyu Lin MEng ’22 wrote the paper alongside CSAIL Research Scientist Andrei Barbu, CSAIL Principal Research Scientist Boris Katz, and MIT-IBM Watson AI Lab Principal Researcher Dan Gutfreund. The researchers are affiliates of the MIT Center for Brains, Minds, and Machines.

      The team is presenting their work at the 2023 Conference on Neural Information Processing Systems (NeurIPS). More

    • 138 Shares169 Views

      in Data Management & Statistics

      Automated system teaches users when to collaborate with an AI assistant

      by

      Artificial intelligence models that pick out patterns in images can often do so better than human eyes — but not always. If a radiologist is using an AI model to help her determine whether a patient’s X-rays show signs of pneumonia, when should she trust the model’s advice and when should she ignore it?

      A customized onboarding process could help this radiologist answer that question, according to researchers at MIT and the MIT-IBM Watson AI Lab. They designed a system that teaches a user when to collaborate with an AI assistant.

      In this case, the training method might find situations where the radiologist trusts the model’s advice — except she shouldn’t because the model is wrong. The system automatically learns rules for how she should collaborate with the AI, and describes them with natural language.

      During onboarding, the radiologist practices collaborating with the AI using training exercises based on these rules, receiving feedback about her performance and the AI’s performance.

      The researchers found that this onboarding procedure led to about a 5 percent improvement in accuracy when humans and AI collaborated on an image prediction task. Their results also show that just telling the user when to trust the AI, without training, led to worse performance.

      Importantly, the researchers’ system is fully automated, so it learns to create the onboarding process based on data from the human and AI performing a specific task. It can also adapt to different tasks, so it can be scaled up and used in many situations where humans and AI models work together, such as in social media content moderation, writing, and programming.

      “So often, people are given these AI tools to use without any training to help them figure out when it is going to be helpful. That’s not what we do with nearly every other tool that people use — there is almost always some kind of tutorial that comes with it. But for AI, this seems to be missing. We are trying to tackle this problem from a methodological and behavioral perspective,” says Hussein Mozannar, a graduate student in the Social and Engineering Systems doctoral program within the Institute for Data, Systems, and Society (IDSS) and lead author of a paper about this training process.

      The researchers envision that such onboarding will be a crucial part of training for medical professionals.

      “One could imagine, for example, that doctors making treatment decisions with the help of AI will first have to do training similar to what we propose. We may need to rethink everything from continuing medical education to the way clinical trials are designed,” says senior author David Sontag, a professor of EECS, a member of the MIT-IBM Watson AI Lab and the MIT Jameel Clinic, and the leader of the Clinical Machine Learning Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      Mozannar, who is also a researcher with the Clinical Machine Learning Group, is joined on the paper by Jimin J. Lee, an undergraduate in electrical engineering and computer science; Dennis Wei, a senior research scientist at IBM Research; and Prasanna Sattigeri and Subhro Das, research staff members at the MIT-IBM Watson AI Lab. The paper will be presented at the Conference on Neural Information Processing Systems.

      Training that evolves

      Existing onboarding methods for human-AI collaboration are often composed of training materials produced by human experts for specific use cases, making them difficult to scale up. Some related techniques rely on explanations, where the AI tells the user its confidence in each decision, but research has shown that explanations are rarely helpful, Mozannar says.

      “The AI model’s capabilities are constantly evolving, so the use cases where the human could potentially benefit from it are growing over time. At the same time, the user’s perception of the model continues changing. So, we need a training procedure that also evolves over time,” he adds.

      To accomplish this, their onboarding method is automatically learned from data. It is built from a dataset that contains many instances of a task, such as detecting the presence of a traffic light from a blurry image.

      The system’s first step is to collect data on the human and AI performing this task. In this case, the human would try to predict, with the help of AI, whether blurry images contain traffic lights.

      The system embeds these data points onto a latent space, which is a representation of data in which similar data points are closer together. It uses an algorithm to discover regions of this space where the human collaborates incorrectly with the AI. These regions capture instances where the human trusted the AI’s prediction but the prediction was wrong, and vice versa.

      Perhaps the human mistakenly trusts the AI when images show a highway at night.

      After discovering the regions, a second algorithm utilizes a large language model to describe each region as a rule, using natural language. The algorithm iteratively fine-tunes that rule by finding contrasting examples. It might describe this region as “ignore AI when it is a highway during the night.”

      These rules are used to build training exercises. The onboarding system shows an example to the human, in this case a blurry highway scene at night, as well as the AI’s prediction, and asks the user if the image shows traffic lights. The user can answer yes, no, or use the AI’s prediction.

      If the human is wrong, they are shown the correct answer and performance statistics for the human and AI on these instances of the task. The system does this for each region, and at the end of the training process, repeats the exercises the human got wrong.

      “After that, the human has learned something about these regions that we hope they will take away in the future to make more accurate predictions,” Mozannar says.

      Onboarding boosts accuracy

      The researchers tested this system with users on two tasks — detecting traffic lights in blurry images and answering multiple choice questions from many domains (such as biology, philosophy, computer science, etc.).

      They first showed users a card with information about the AI model, how it was trained, and a breakdown of its performance on broad categories. Users were split into five groups: Some were only shown the card, some went through the researchers’ onboarding procedure, some went through a baseline onboarding procedure, some went through the researchers’ onboarding procedure and were given recommendations of when they should or should not trust the AI, and others were only given the recommendations.

      Only the researchers’ onboarding procedure without recommendations improved users’ accuracy significantly, boosting their performance on the traffic light prediction task by about 5 percent without slowing them down. However, onboarding was not as effective for the question-answering task. The researchers believe this is because the AI model, ChatGPT, provided explanations with each answer that convey whether it should be trusted.

      But providing recommendations without onboarding had the opposite effect — users not only performed worse, they took more time to make predictions.

      “When you only give someone recommendations, it seems like they get confused and don’t know what to do. It derails their process. People also don’t like being told what to do, so that is a factor as well,” Mozannar says.

      Providing recommendations alone could harm the user if those recommendations are wrong, he adds. With onboarding, on the other hand, the biggest limitation is the amount of available data. If there aren’t enough data, the onboarding stage won’t be as effective, he says.

      In the future, he and his collaborators want to conduct larger studies to evaluate the short- and long-term effects of onboarding. They also want to leverage unlabeled data for the onboarding process, and find methods to effectively reduce the number of regions without omitting important examples.

      “People are adopting AI systems willy-nilly, and indeed AI offers great potential, but these AI agents still sometimes makes mistakes. Thus, it’s crucial for AI developers to devise methods that help humans know when it’s safe to rely on the AI’s suggestions,” says Dan Weld, professor emeritus at the Paul G. Allen School of Computer Science and Engineering at the University of Washington, who was not involved with this research. “Mozannar et al. have created an innovative method for identifying situations where the AI is trustworthy, and (importantly) to describe them to people in a way that leads to better human-AI team interactions.”

      This work is funded, in part, by the MIT-IBM Watson AI Lab. More

    • 113 Shares119 Views

      in Data Management & Statistics

      Synthetic imagery sets new bar in AI training efficiency

      by

      Data is the new soil, and in this fertile new ground, MIT researchers are planting more than just pixels. By using synthetic images to train machine learning models, a team of scientists recently surpassed results obtained from traditional “real-image” training methods. 

      At the core of the approach is a system called StableRep, which doesn’t just use any synthetic images; it generates them through ultra-popular text-to-image models like Stable Diffusion. It’s like creating worlds with words. 

      So what’s in StableRep’s secret sauce? A strategy called “multi-positive contrastive learning.”

      “We’re teaching the model to learn more about high-level concepts through context and variance, not just feeding it data,” says Lijie Fan, MIT PhD student in electrical engineering, affiliate of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), lead researcher on the work. “When multiple images, all generated from the same text, all treated as depictions of the same underlying thing, the model dives deeper into the concepts behind the images, say the object, not just their pixels.”

      This approach considers multiple images spawned from identical text prompts as positive pairs, providing additional information during training, not just adding more diversity but specifying to the vision system which images are alike and which are different. Remarkably, StableRep outshone the prowess of top-tier models trained on real images, such as SimCLR and CLIP, in extensive datasets.

      “While StableRep helps mitigate the challenges of data acquisition in machine learning, it also ushers in a stride towards a new era of AI training techniques. The capacity to produce high-caliber, diverse synthetic images on command could help curtail cumbersome expenses and resources,” says Fan. 

      The process of data collection has never been straightforward. Back in the 1990s, researchers had to manually capture photographs to assemble datasets for objects and faces. The 2000s saw individuals scouring the internet for data. However, this raw, uncurated data often contained discrepancies when compared to real-world scenarios and reflected societal biases, presenting a distorted view of reality. The task of cleansing datasets through human intervention is not only expensive, but also exceedingly challenging. Imagine, though, if this arduous data collection could be distilled down to something as simple as issuing a command in natural language. 

      A pivotal aspect of StableRep’s triumph is the adjustment of the “guidance scale” in the generative model, which ensures a delicate balance between the synthetic images’ diversity and fidelity. When finely tuned, synthetic images used in training these self-supervised models were found to be as effective, if not more so, than real images.

      Taking it a step forward, language supervision was added to the mix, creating an enhanced variant: StableRep+. When trained with 20 million synthetic images, StableRep+ not only achieved superior accuracy but also displayed remarkable efficiency compared to CLIP models trained with a staggering 50 million real images.

      Yet, the path ahead isn’t without its potholes. The researchers candidly address several limitations, including the current slow pace of image generation, semantic mismatches between text prompts and the resultant images, potential amplification of biases, and complexities in image attribution, all of which are imperative to address for future advancements. Another issue is that StableRep requires first training the generative model on large-scale real data. The team acknowledges that starting with real data remains a necessity; however, when you have a good generative model, you can repurpose it for new tasks, like training recognition models and visual representations. 

      The team notes that they haven’t gotten around the need to start with real data; it’s just that once you have a good generative model you can repurpose it for new tasks, like training recognition models and visual representations. 

      While StableRep offers a good solution by diminishing the dependency on vast real-image collections, it brings to the fore concerns regarding hidden biases within the uncurated data used for these text-to-image models. The choice of text prompts, integral to the image synthesis process, is not entirely free from bias, “indicating the essential role of meticulous text selection or possible human curation,” says Fan. 

      “Using the latest text-to-image models, we’ve gained unprecedented control over image generation, allowing for a diverse range of visuals from a single text input. This surpasses real-world image collection in efficiency and versatility. It proves especially useful in specialized tasks, like balancing image variety in long-tail recognition, presenting a practical supplement to using real images for training,” says Fan. “Our work signifies a step forward in visual learning, towards the goal of offering cost-effective training alternatives while highlighting the need for ongoing improvements in data quality and synthesis.”

      “One dream of generative model learning has long been to be able to generate data useful for discriminative model training,” says Google DeepMind researcher and University of Toronto professor of computer science David Fleet, who was not involved in the paper. “While we have seen some signs of life, the dream has been elusive, especially on large-scale complex domains like high-resolution images. This paper provides compelling evidence, for the first time to my knowledge, that the dream is becoming a reality. They show that contrastive learning from massive amounts of synthetic image data can produce representations that outperform those learned from real data at scale, with the potential to improve myriad downstream vision tasks.”

      Fan is joined by Yonglong Tian PhD ’22 as lead authors of the paper, as well as MIT associate professor of electrical engineering and computer science and CSAIL principal investigator Phillip Isola; Google researcher and OpenAI technical staff member Huiwen Chang; and Google staff research scientist Dilip Krishnan. The team will present StableRep at the 2023 Conference on Neural Information Processing Systems (NeurIPS) in New Orleans. More

    • 88 Shares129 Views

      in Data Management & Statistics

      Accelerating AI tasks while preserving data security

      by

      With the proliferation of computationally intensive machine-learning applications, such as chatbots that perform real-time language translation, device manufacturers often incorporate specialized hardware components to rapidly move and process the massive amounts of data these systems demand.

      Choosing the best design for these components, known as deep neural network accelerators, is challenging because they can have an enormous range of design options. This difficult problem becomes even thornier when a designer seeks to add cryptographic operations to keep data safe from attackers.

      Now, MIT researchers have developed a search engine that can efficiently identify optimal designs for deep neural network accelerators, that preserve data security while boosting performance.

      Their search tool, known as SecureLoop, is designed to consider how the addition of data encryption and authentication measures will impact the performance and energy usage of the accelerator chip. An engineer could use this tool to obtain the optimal design of an accelerator tailored to their neural network and machine-learning task.

      When compared to conventional scheduling techniques that don’t consider security, SecureLoop can improve performance of accelerator designs while keeping data protected.  

      Using SecureLoop could help a user improve the speed and performance of demanding AI applications, such as autonomous driving or medical image classification, while ensuring sensitive user data remains safe from some types of attacks.

      “If you are interested in doing a computation where you are going to preserve the security of the data, the rules that we used before for finding the optimal design are now broken. So all of that optimization needs to be customized for this new, more complicated set of constraints. And that is what [lead author] Kyungmi has done in this paper,” says Joel Emer, an MIT professor of the practice in computer science and electrical engineering and co-author of a paper on SecureLoop.

      Emer is joined on the paper by lead author Kyungmi Lee, an electrical engineering and computer science graduate student; Mengjia Yan, the Homer A. Burnell Career Development Assistant Professor of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Anantha Chandrakasan, dean of the MIT School of Engineering and the Vannevar Bush Professor of Electrical Engineering and Computer Science. The research will be presented at the IEEE/ACM International Symposium on Microarchitecture.

      “The community passively accepted that adding cryptographic operations to an accelerator will introduce overhead. They thought it would introduce only a small variance in the design trade-off space. But, this is a misconception. In fact, cryptographic operations can significantly distort the design space of energy-efficient accelerators. Kyungmi did a fantastic job identifying this issue,” Yan adds.

      Secure acceleration

      A deep neural network consists of many layers of interconnected nodes that process data. Typically, the output of one layer becomes the input of the next layer. Data are grouped into units called tiles for processing and transfer between off-chip memory and the accelerator. Each layer of the neural network can have its own data tiling configuration.

      A deep neural network accelerator is a processor with an array of computational units that parallelizes operations, like multiplication, in each layer of the network. The accelerator schedule describes how data are moved and processed.

      Since space on an accelerator chip is at a premium, most data are stored in off-chip memory and fetched by the accelerator when needed. But because data are stored off-chip, they are vulnerable to an attacker who could steal information or change some values, causing the neural network to malfunction.

      “As a chip manufacturer, you can’t guarantee the security of external devices or the overall operating system,” Lee explains.

      Manufacturers can protect data by adding authenticated encryption to the accelerator. Encryption scrambles the data using a secret key. Then authentication cuts the data into uniform chunks and assigns a cryptographic hash to each chunk of data, which is stored along with the data chunk in off-chip memory.

      When the accelerator fetches an encrypted chunk of data, known as an authentication block, it uses a secret key to recover and verify the original data before processing it.

      But the sizes of authentication blocks and tiles of data don’t match up, so there could be multiple tiles in one block, or a tile could be split between two blocks. The accelerator can’t arbitrarily grab a fraction of an authentication block, so it may end up grabbing extra data, which uses additional energy and slows down computation.

      Plus, the accelerator still must run the cryptographic operation on each authentication block, adding even more computational cost.

      An efficient search engine

      With SecureLoop, the MIT researchers sought a method that could identify the fastest and most energy efficient accelerator schedule — one that minimizes the number of times the device needs to access off-chip memory to grab extra blocks of data because of encryption and authentication.  

      They began by augmenting an existing search engine Emer and his collaborators previously developed, called Timeloop. First, they added a model that could account for the additional computation needed for encryption and authentication.

      Then, they reformulated the search problem into a simple mathematical expression, which enables SecureLoop to find the ideal authentical block size in a much more efficient manner than searching through all possible options.

      “Depending on how you assign this block, the amount of unnecessary traffic might increase or decrease. If you assign the cryptographic block cleverly, then you can just fetch a small amount of additional data,” Lee says.

      Finally, they incorporated a heuristic technique that ensures SecureLoop identifies a schedule which maximizes the performance of the entire deep neural network, rather than only a single layer.

      At the end, the search engine outputs an accelerator schedule, which includes the data tiling strategy and the size of the authentication blocks, that provides the best possible speed and energy efficiency for a specific neural network.

      “The design spaces for these accelerators are huge. What Kyungmi did was figure out some very pragmatic ways to make that search tractable so she could find good solutions without needing to exhaustively search the space,” says Emer.

      When tested in a simulator, SecureLoop identified schedules that were up to 33.2 percent faster and exhibited 50.2 percent better energy delay product (a metric related to energy efficiency) than other methods that didn’t consider security.

      The researchers also used SecureLoop to explore how the design space for accelerators changes when security is considered. They learned that allocating a bit more of the chip’s area for the cryptographic engine and sacrificing some space for on-chip memory can lead to better performance, Lee says.

      In the future, the researchers want to use SecureLoop to find accelerator designs that are resilient to side-channel attacks, which occur when an attacker has access to physical hardware. For instance, an attacker could monitor the power consumption pattern of a device to obtain secret information, even if the data have been encrypted. They are also extending SecureLoop so it could be applied to other kinds of computation.

      This work is funded, in part, by Samsung Electronics and the Korea Foundation for Advanced Studies. More

    • 125 Shares199 Views

      in Data Management & Statistics

      New techniques efficiently accelerate sparse tensors for massive AI models

      by

      Researchers from MIT and NVIDIA have developed two techniques that accelerate the processing of sparse tensors, a type of data structure that’s used for high-performance computing tasks. The complementary techniques could result in significant improvements to the performance and energy-efficiency of systems like the massive machine-learning models that drive generative artificial intelligence.

      Tensors are data structures used by machine-learning models. Both of the new methods seek to efficiently exploit what’s known as sparsity — zero values — in the tensors. When processing these tensors, one can skip over the zeros and save on both computation and memory. For instance, anything multiplied by zero is zero, so it can skip that operation. And it can compress the tensor (zeros don’t need to be stored) so a larger portion can be stored in on-chip memory.

      However, there are several challenges to exploiting sparsity. Finding the nonzero values in a large tensor is no easy task. Existing approaches often limit the locations of nonzero values by enforcing a sparsity pattern to simplify the search, but this limits the variety of sparse tensors that can be processed efficiently.

      Another challenge is that the number of nonzero values can vary in different regions of the tensor. This makes it difficult to determine how much space is required to store different regions in memory. To make sure the region fits, more space is often allocated than is needed, causing the storage buffer to be underutilized. This increases off-chip memory traffic, which increases energy consumption.

      The MIT and NVIDIA researchers crafted two solutions to address these problems. For one, they developed a technique that allows the hardware to efficiently find the nonzero values for a wider variety of sparsity patterns.

      For the other solution, they created a method that can handle the case where the data do not fit in memory, which increases the utilization of the storage buffer and reduces off-chip memory traffic.

      Both methods boost the performance and reduce the energy demands of hardware accelerators specifically designed to speed up the processing of sparse tensors.

      “Typically, when you use more specialized or domain-specific hardware accelerators, you lose the flexibility that you would get from a more general-purpose processor, like a CPU. What stands out with these two works is that we show that you can still maintain flexibility and adaptability while being specialized and efficient,” says Vivienne Sze, associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS), a member of the Research Laboratory of Electronics (RLE), and co-senior author of papers on both advances.

      Her co-authors include lead authors Yannan Nellie Wu PhD ’23 and Zi Yu Xue, an electrical engineering and computer science graduate student; and co-senior author Joel Emer, an MIT professor of the practice in computer science and electrical engineering and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL), as well as others at NVIDIA. Both papers will be presented at the IEEE/ACM International Symposium on Microarchitecture.

      HighLight: Efficiently finding zero values

      Sparsity can arise in the tensor for a variety of reasons. For example, researchers sometimes “prune” unnecessary pieces of the machine-learning models by replacing some values in the tensor with zeros, creating sparsity. The degree of sparsity (percentage of zeros) and the locations of the zeros can vary for different models.

      To make it easier to find the remaining nonzero values in a model with billions of individual values, researchers often restrict the location of the nonzero values so they fall into a certain pattern. However, each hardware accelerator is typically designed to support one specific sparsity pattern, limiting its flexibility.  

      By contrast, the hardware accelerator the MIT researchers designed, called HighLight, can handle a wide variety of sparsity patterns and still perform well when running models that don’t have any zero values.

      They use a technique they call “hierarchical structured sparsity” to efficiently represent a wide variety of sparsity patterns that are composed of several simple sparsity patterns. This approach divides the values in a tensor into smaller blocks, where each block has its own simple, sparsity pattern (perhaps two zeros and two nonzeros in a block with four values).

      Then, they combine the blocks into a hierarchy, where each collection of blocks also has its own simple, sparsity pattern (perhaps one zero block and three nonzero blocks in a level with four blocks). They continue combining blocks into larger levels, but the patterns remain simple at each step.

      This simplicity enables HighLight to more efficiently find and skip zeros, so it can take full advantage of the opportunity to cut excess computation. On average, their accelerator design had about six times better energy-delay product (a metric related to energy efficiency) than other approaches.

      “In the end, the HighLight accelerator is able to efficiently accelerate dense models because it does not introduce a lot of overhead, and at the same time it is able to exploit workloads with different amounts of zero values based on hierarchical structured sparsity,” Wu explains.

      In the future, she and her collaborators want to apply hierarchical structured sparsity to more types of machine-learning models and different types of tensors in the models.

      Tailors and Swiftiles: Effectively “overbooking” to accelerate workloads

      Researchers can also leverage sparsity to more efficiently move and process data on a computer chip.

      Since the tensors are often larger than what can be stored in the memory buffer on chip, the chip only grabs and processes a chunk of the tensor at a time. The chunks are called tiles.

      To maximize the utilization of that buffer and limit the number of times the chip must access off-chip memory, which often dominates energy consumption and limits processing speed, researchers seek to use the largest tile that will fit into the buffer.

      But in a sparse tensor, many of the data values are zero, so an even larger tile can fit into the buffer than one might expect based on its capacity. Zero values don’t need to be stored.

      But the number of zero values can vary across different regions of the tensor, so they can also vary for each tile. This makes it difficult to determine a tile size that will fit in the buffer. As a result, existing approaches often conservatively assume there are no zeros and end up selecting a smaller tile, which results in wasted blank spaces in the buffer.

      To address this uncertainty, the researchers propose the use of “overbooking” to allow them to increase the tile size, as well as a way to tolerate it if the tile doesn’t fit the buffer.

      The same way an airline overbooks tickets for a flight, if all the passengers show up, the airline must compensate the ones who are bumped from the plane. But usually all the passengers don’t show up.

      In a sparse tensor, a tile size can be chosen such that usually the tiles will have enough zeros that most still fit into the buffer. But occasionally, a tile will have more nonzero values than will fit. In this case, those data are bumped out of the buffer.

      The researchers enable the hardware to only re-fetch the bumped data without grabbing and processing the entire tile again. They modify the “tail end” of the buffer to handle this, hence the name of this technique, Tailors.

      Then they also created an approach for finding the size for tiles that takes advantage of overbooking. This method, called Swiftiles, swiftly estimates the ideal tile size so that a specific percentage of tiles, set by the user, are overbooked. (The names “Tailors” and “Swiftiles” pay homage to Taylor Swift, whose recent Eras tour was fraught with overbooked presale codes for tickets).

      Swiftiles reduces the number of times the hardware needs to check the tensor to identify an ideal tile size, saving on computation. The combination of Tailors and Swiftiles more than doubles the speed while requiring only half the energy demands of existing hardware accelerators which cannot handle overbooking.

      “Swiftiles allows us to estimate how large these tiles need to be without requiring multiple iterations to refine the estimate. This only works because overbooking is supported. Even if you are off by a decent amount, you can still extract a fair bit of speedup because of the way the non-zeros are distributed,” Xue says.

      In the future, the researchers want to apply the idea of overbooking to other aspects in computer architecture and also work to improve the process for estimating the optimal level of overbooking.

      This research is funded, in part, by the MIT AI Hardware Program. More

    • 75 Shares149 Views

      in Data Management & Statistics

      Making genetic prediction models more inclusive

      by

      While any two human genomes are about 99.9 percent identical, genetic variation in the remaining 0.1 percent plays an important role in shaping human diversity, including a person’s risk for developing certain diseases.

      Measuring the cumulative effect of these small genetic differences can provide an estimate of an individual’s genetic risk for a particular disease or their likelihood of having a particular trait. However, the majority of models used to generate these “polygenic scores” are based on studies done in people of European descent, and do not accurately gauge the risk for people of non-European ancestry or people whose genomes contain a mixture of chromosome regions inherited from previously isolated populations, also known as admixed ancestry.

      In an effort to make these genetic scores more inclusive, MIT researchers have created a new model that takes into account genetic information from people from a wider diversity of genetic ancestries across the world. Using this model, they showed that they could increase the accuracy of genetics-based predictions for a variety of traits, especially for people from populations that have been traditionally underrepresented in genetic studies.

      “For people of African ancestry, our model proved to be about 60 percent more accurate on average,” says Manolis Kellis, a professor of computer science in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and a member of the Broad Institute of MIT and Harvard. “For people of admixed genetic backgrounds more broadly, who have been excluded from most previous models, the accuracy of our model increased by an average of about 18 percent.”

      The researchers hope their more inclusive modeling approach could help improve health outcomes for a wider range of people and promote health equity by spreading the benefits of genomic sequencing more widely across the globe.

      “What we have done is created a method that allows you to be much more accurate for admixed and ancestry-diverse individuals, and ensure the results and the benefits of human genetics research are equally shared by everyone,” says MIT postdoc Yosuke Tanigawa, the lead and co-corresponding author of the paper, which appears today in open-access form in the American Journal of Human Genetics. The researchers have made all of their data publicly available for the broader scientific community to use.

      More inclusive models

      The work builds on the Human Genome Project, which mapped all of the genes found in the human genome, and on subsequent large-scale, cohort-based studies of how genetic variants in the human genome are linked to disease risk and other differences between individuals.

      These studies showed that the effect of any individual genetic variant on its own is typically very small. Together, these small effects add up and influence the risk of developing heart disease or diabetes, having a stroke, or being diagnosed with psychiatric disorders such as schizophrenia.

      “We have hundreds of thousands of genetic variants that are associated with complex traits, each of which is individually playing a weak effect, but together they are beginning to be predictive for disease predispositions,” Kellis says.

      However, most of these genome-wide association studies included few people of non-European descent, so polygenic risk models based on them translate poorly to non-European populations. People from different geographic areas can have different patterns of genetic variation, shaped by stochastic drift, population history, and environmental factors — for example, in people of African descent, genetic variants that protect against malaria are more common than in other populations. Those variants also affect other traits involving the immune system, such as counts of neutrophils, a type of immune cell. That variation would not be well-captured in a model based on genetic analysis of people of European ancestry alone.

      “If you are an individual of African descent, of Latin American descent, of Asian descent, then you are currently being left out by the system,” Kellis says. “This inequity in the utilization of genetic information for predicting risk of patients can cause unnecessary burden, unnecessary deaths, and unnecessary lack of prevention, and that’s where our work comes in.”

      Some researchers have begun trying to address these disparities by creating distinct models for people of European descent, of African descent, or of Asian descent. These emerging approaches assign individuals to distinct genetic ancestry groups, aggregate the data to create an association summary, and make genetic prediction models. However, these approaches still don’t represent people of admixed genetic backgrounds well.

      “Our approach builds on the previous work without requiring researchers to assign individuals or local genomic segments of individuals to predefined distinct genetic ancestry groups,” Tanigawa says. “Instead, we develop a single model for everybody by directly working on individuals across the continuum of their genetic ancestries.”

      In creating their new model, the MIT team used computational and statistical techniques that enabled them to study each individual’s unique genetic profile instead of grouping individuals by population. This methodological advancement allowed the researchers to include people of admixed ancestry, who made up nearly 10 percent of the UK Biobank dataset used for this study and currently account for about one in seven newborns in the United States.

      “Because we work at the individual level, there is no need for computing summary-level data for different populations,” Kellis says. “Thus, we did not need to exclude individuals of admixed ancestry, increasing our power by including more individuals and representing contributions from all populations in our combined model.”

      Better predictions

      To create their new model, the researchers used genetic data from more than 280,000 people, which was collected by UK Biobank, a large-scale biomedical database and research resource containing de-identified genetic, lifestyle, and health information from half a million U.K. participants. Using another set of about 81,000 held-out individuals from the UK Biobank, the researchers evaluated their model across 60 traits, which included traits related to body size and shape, such as height and body mass index, as well as blood traits such as white blood cell count and red blood cell count, which also have a genetic basis.

      The researchers found that, compared to models trained only on European-ancestry individuals, their model’s predictions are more accurate for all genetic ancestry groups. The most notable gain was for people of African ancestry, who showed 61 percent average improvements, even though they only made up about 1.5 percent of samples in UK Biobank. The researchers also saw improvements of 11 percent for people of South Asian descent and 5 percent for white British people. Predictions for people of admixed ancestry improved by about 18 percent.

      “When you bring all the individuals together in the training set, everybody contributes to the training of the polygenic score modeling on equal footing,” Tanigawa says. “Combined with increasingly more inclusive data collection efforts, our method can help leverage these efforts to improve predictive accuracy for all.”

      The MIT team hopes its approach can eventually be incorporated into tests of an individual’s risk of a variety of diseases. Such tests could be combined with conventional risk factors and used to help doctors diagnose disease or to help people manage their risk for certain diseases before they develop.

      “Our work highlights the power of diversity, equity, and inclusion efforts in the context of genomics research,” Tanigawa says.

      The researchers now hope to add even more data to their model, including data from the United States, and to apply it to additional traits that they didn’t analyze in this study.

      “This is just the start,” Kellis says. “We can’t wait to see more people join our effort to propel inclusive human genetics research.”

      The research was funded by the National Institutes of Health. More