Algorithms

Subterms

    More stories

    • 100 Shares109 Views

      in Data Management & Statistics

      Researchers release open-source photorealistic simulator for autonomous driving

      by

      Hyper-realistic virtual worlds have been heralded as the best driving schools for autonomous vehicles (AVs), since they’ve proven fruitful test beds for safely trying out dangerous driving scenarios. Tesla, Waymo, and other self-driving companies all rely heavily on data to enable expensive and proprietary photorealistic simulators, since testing and gathering nuanced I-almost-crashed data usually isn’t the most easy or desirable to recreate. 

      To that end, scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) created “VISTA 2.0,” a data-driven simulation engine where vehicles can learn to drive in the real world and recover from near-crash scenarios. What’s more, all of the code is being open-sourced to the public. 

      “Today, only companies have software like the type of simulation environments and capabilities of VISTA 2.0, and this software is proprietary. With this release, the research community will have access to a powerful new tool for accelerating the research and development of adaptive robust control for autonomous driving,” says MIT Professor and CSAIL Director Daniela Rus, senior author on a paper about the research. 

      Play video

      VISTA is a data-driven, photorealistic simulator for autonomous driving. It can simulate not just live video but LiDAR data and event cameras, and also incorporate other simulated vehicles to model complex driving situations. VISTA is open source and the code can be found below.

      VISTA 2.0 builds off of the team’s previous model, VISTA, and it’s fundamentally different from existing AV simulators since it’s data-driven — meaning it was built and photorealistically rendered from real-world data — thereby enabling direct transfer to reality. While the initial iteration supported only single car lane-following with one camera sensor, achieving high-fidelity data-driven simulation required rethinking the foundations of how different sensors and behavioral interactions can be synthesized. 

      Enter VISTA 2.0: a data-driven system that can simulate complex sensor types and massively interactive scenarios and intersections at scale. With much less data than previous models, the team was able to train autonomous vehicles that could be substantially more robust than those trained on large amounts of real-world data. 

      “This is a massive jump in capabilities of data-driven simulation for autonomous vehicles, as well as the increase of scale and ability to handle greater driving complexity,” says Alexander Amini, CSAIL PhD student and co-lead author on two new papers, together with fellow PhD student Tsun-Hsuan Wang. “VISTA 2.0 demonstrates the ability to simulate sensor data far beyond 2D RGB cameras, but also extremely high dimensional 3D lidars with millions of points, irregularly timed event-based cameras, and even interactive and dynamic scenarios with other vehicles as well.” 

      The team was able to scale the complexity of the interactive driving tasks for things like overtaking, following, and negotiating, including multiagent scenarios in highly photorealistic environments. 

      Training AI models for autonomous vehicles involves hard-to-secure fodder of different varieties of edge cases and strange, dangerous scenarios, because most of our data (thankfully) is just run-of-the-mill, day-to-day driving. Logically, we can’t just crash into other cars just to teach a neural network how to not crash into other cars.

      Recently, there’s been a shift away from more classic, human-designed simulation environments to those built up from real-world data. The latter have immense photorealism, but the former can easily model virtual cameras and lidars. With this paradigm shift, a key question has emerged: Can the richness and complexity of all of the sensors that autonomous vehicles need, such as lidar and event-based cameras that are more sparse, accurately be synthesized? 

      Lidar sensor data is much harder to interpret in a data-driven world — you’re effectively trying to generate brand-new 3D point clouds with millions of points, only from sparse views of the world. To synthesize 3D lidar point clouds, the team used the data that the car collected, projected it into a 3D space coming from the lidar data, and then let a new virtual vehicle drive around locally from where that original vehicle was. Finally, they projected all of that sensory information back into the frame of view of this new virtual vehicle, with the help of neural networks. 

      Together with the simulation of event-based cameras, which operate at speeds greater than thousands of events per second, the simulator was capable of not only simulating this multimodal information, but also doing so all in real time — making it possible to train neural nets offline, but also test online on the car in augmented reality setups for safe evaluations. “The question of if multisensor simulation at this scale of complexity and photorealism was possible in the realm of data-driven simulation was very much an open question,” says Amini. 

      With that, the driving school becomes a party. In the simulation, you can move around, have different types of controllers, simulate different types of events, create interactive scenarios, and just drop in brand new vehicles that weren’t even in the original data. They tested for lane following, lane turning, car following, and more dicey scenarios like static and dynamic overtaking (seeing obstacles and moving around so you don’t collide). With the multi-agency, both real and simulated agents interact, and new agents can be dropped into the scene and controlled any which way. 

      Taking their full-scale car out into the “wild” — a.k.a. Devens, Massachusetts — the team saw  immediate transferability of results, with both failures and successes. They were also able to demonstrate the bodacious, magic word of self-driving car models: “robust.” They showed that AVs, trained entirely in VISTA 2.0, were so robust in the real world that they could handle that elusive tail of challenging failures. 

      Now, one guardrail humans rely on that can’t yet be simulated is human emotion. It’s the friendly wave, nod, or blinker switch of acknowledgement, which are the type of nuances the team wants to implement in future work. 

      “The central algorithm of this research is how we can take a dataset and build a completely synthetic world for learning and autonomy,” says Amini. “It’s a platform that I believe one day could extend in many different axes across robotics. Not just autonomous driving, but many areas that rely on vision and complex behaviors. We’re excited to release VISTA 2.0 to help enable the community to collect their own datasets and convert them into virtual worlds where they can directly simulate their own virtual autonomous vehicles, drive around these virtual terrains, train autonomous vehicles in these worlds, and then can directly transfer them to full-sized, real self-driving cars.” 

      Amini and Wang wrote the paper alongside Zhijian Liu, MIT CSAIL PhD student; Igor Gilitschenski, assistant professor in computer science at the University of Toronto; Wilko Schwarting, AI research scientist and MIT CSAIL PhD ’20; Song Han, associate professor at MIT’s Department of Electrical Engineering and Computer Science; Sertac Karaman, associate professor of aeronautics and astronautics at MIT; and Daniela Rus, MIT professor and CSAIL director. The researchers presented the work at the IEEE International Conference on Robotics and Automation (ICRA) in Philadelphia. 

      This work was supported by the National Science Foundation and Toyota Research Institute. The team acknowledges the support of NVIDIA with the donation of the Drive AGX Pegasus. More

    • 63 Shares189 Views

      in Data Management & Statistics

      Hallucinating to better text translation

      by

      As babies, we babble and imitate our way to learning languages. We don’t start off reading raw text, which requires fundamental knowledge and understanding about the world, as well as the advanced ability to interpret and infer descriptions and relationships. Rather, humans begin our language journey slowly, by pointing and interacting with our environment, basing our words and perceiving their meaning through the context of the physical and social world. Eventually, we can craft full sentences to communicate complex ideas.

      Similarly, when humans begin learning and translating into another language, the incorporation of other sensory information, like multimedia, paired with the new and unfamiliar words, like flashcards with images, improves language acquisition and retention. Then, with enough practice, humans can accurately translate new, unseen sentences in context without the accompanying media; however, imagining a picture based on the original text helps.

      This is the basis of a new machine learning model, called VALHALLA, by researchers from MIT, IBM, and the University of California at San Diego, in which a trained neural network sees a source sentence in one language, hallucinates an image of what it looks like, and then uses both to translate into a target language. The team found that their method demonstrates improved accuracy of machine translation over text-only translation. Further, it provided an additional boost for cases with long sentences, under-resourced languages, and instances where part of the source sentence is inaccessible to the machine translator.

      As a core task within the AI field of natural language processing (NLP), machine translation is an “eminently practical technology that’s being used by millions of people every day,” says study co-author Yoon Kim, assistant professor in MIT’s Department of Electrical Engineering and Computer Science with affiliations in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT-IBM Watson AI Lab. With recent, significant advances in deep learning, “there’s been an interesting development in how one might use non-text information — for example, images, audio, or other grounding information — to tackle practical tasks involving language” says Kim, because “when humans are performing language processing tasks, we’re doing so within a grounded, situated world.” The pairing of hallucinated images and text during inference, the team postulated, imitates that process, providing context for improved performance over current state-of-the-art techniques, which utilize text-only data.

      This research will be presented at the IEEE / CVF Computer Vision and Pattern Recognition Conference this month. Kim’s co-authors are UC San Diego graduate student Yi Li and Professor Nuno Vasconcelos, along with research staff members Rameswar Panda, Chun-fu “Richard” Chen, Rogerio Feris, and IBM Director David Cox of IBM Research and the MIT-IBM Watson AI Lab.

      Learning to hallucinate from images

      When we learn new languages and to translate, we’re often provided with examples and practice before venturing out on our own. The same is true for machine-translation systems; however, if images are used during training, these AI methods also require visual aids for testing, limiting their applicability, says Panda.

      “In real-world scenarios, you might not have an image with respect to the source sentence. So, our motivation was basically: Instead of using an external image during inference as input, can we use visual hallucination — the ability to imagine visual scenes — to improve machine translation systems?” says Panda.

      To do this, the team used an encoder-decoder architecture with two transformers, a type of neural network model that’s suited for sequence-dependent data, like language, that can pay attention key words and semantics of a sentence. One transformer generates a visual hallucination, and the other performs multimodal translation using outputs from the first transformer.

      During training, there are two streams of translation: a source sentence and a ground-truth image that is paired with it, and the same source sentence that is visually hallucinated to make a text-image pair. First the ground-truth image and sentence are tokenized into representations that can be handled by transformers; for the case of the sentence, each word is a token. The source sentence is tokenized again, but this time passed through the visual hallucination transformer, outputting a hallucination, a discrete image representation of the sentence. The researchers incorporated an autoregression that compares the ground-truth and hallucinated representations for congruency — e.g., homonyms: a reference to an animal “bat” isn’t hallucinated as a baseball bat. The hallucination transformer then uses the difference between them to optimize its predictions and visual output, making sure the context is consistent.

      The two sets of tokens are then simultaneously passed through the multimodal translation transformer, each containing the sentence representation and either the hallucinated or ground-truth image. The tokenized text translation outputs are compared with the goal of being similar to each other and to the target sentence in another language. Any differences are then relayed back to the translation transformer for further optimization.

      For testing, the ground-truth image stream drops off, since images likely wouldn’t be available in everyday scenarios.

      “To the best of our knowledge, we haven’t seen any work which actually uses a hallucination transformer jointly with a multimodal translation system to improve machine translation performance,” says Panda.

      Visualizing the target text

      To test their method, the team put VALHALLA up against other state-of-the-art multimodal and text-only translation methods. They used public benchmark datasets containing ground-truth images with source sentences, and a dataset for translating text-only news articles. The researchers measured its performance over 13 tasks, ranging from translation on well-resourced languages (like English, German, and French), under-resourced languages (like English to Romanian) and non-English (like Spanish to French). The group also tested varying transformer model sizes, how accuracy changes with the sentence length, and translation under limited textual context, where portions of the text were hidden from the machine translators.

      The team observed significant improvements over text-only translation methods, improving data efficiency, and that smaller models performed better than the larger base model. As sentences became longer, VALHALLA’s performance over other methods grew, which the researchers attributed to the addition of more ambiguous words. In cases where part of the sentence was masked, VALHALLA could recover and translate the original text, which the team found surprising.

      Further unexpected findings arose: “Where there weren’t as many training [image and] text pairs, [like for under-resourced languages], improvements were more significant, which indicates that grounding in images helps in low-data regimes,” says Kim. “Another thing that was quite surprising to me was this improved performance, even on types of text that aren’t necessarily easily connectable to images. For example, maybe it’s not so surprising if this helps in translating visually salient sentences, like the ‘there is a red car in front of the house.’ [However], even in text-only [news article] domains, the approach was able to improve upon text-only systems.”

      While VALHALLA performs well, the researchers note that it does have limitations, requiring pairs of sentences to be annotated with an image, which could make it more expensive to obtain. It also performs better in its ground domain and not the text-only news articles. Moreover, Kim and Panda note, a technique like VALHALLA is still a black box, with the assumption that hallucinated images are providing helpful information, and the team plans to investigate what and how the model is learning in order to validate their methods.

      In the future, the team plans to explore other means of improving translation. “Here, we only focus on images, but there are other types of a multimodal information — for example, speech, video or touch, or other sensory modalities,” says Panda. “We believe such multimodal grounding can lead to even more efficient machine translation models, potentially benefiting translation across many low-resource languages spoken in the world.”

      This research was supported, in part, by the MIT-IBM Watson AI Lab and the National Science Foundation. More

    • 125 Shares189 Views

      in Data Management & Statistics

      Emery Brown wins a share of 2022 Gruber Neuroscience Prize

      by

      The Gruber Foundation announced on May 17 that Emery N. Brown, the Edward Hood Taplin Professor of Medical Engineering and Computational Neuroscience at MIT, has won the 2022 Gruber Neuroscience Prize along with neurophysicists Laurence Abbott of Columbia University, Terrence Sejnowski of the Salk Institute for Biological Studies, and Haim Sompolinsky of the Hebrew University of Jerusalem.

      The foundation says it honored the four recipients for their influential contributions to the fields of computational and theoretical neuroscience. As datasets have grown ever larger and more complex, these fields have increasingly helped scientists unravel the mysteries of how the brain functions in both health and disease. The prize, which includes a total $500,000 award, will be presented in San Diego, California, on Nov. 13 at the annual meeting of the Society for Neuroscience.

      “These four remarkable scientists have applied their expertise in mathematical and statistical analysis, physics, and machine learning to create theories, mathematical models, and tools that have greatly advanced how we study and understand the brain,” says Joshua Sanes, professor of molecular and cellular biology and founding director of the Center for Brain Science at Harvard University and member of the selection advisory board to the prize. “Their insights and research have not only transformed how experimental neuroscientists do their research, but also are leading to promising new ways of providing clinical care.”

      Brown, who is an investigator in The Picower Institute for Learning and Memory and the Institute for Medical Engineering and Science at MIT, an anesthesiologist at Massachusetts General Hospital, and a professor at Harvard Medical School, says: “It is a pleasant surprise and tremendous honor to be named a co-recipient of the 2022 Gruber Prize in Neuroscience. I am especially honored to share this award with three luminaries in computational and theoretical neuroscience.”

      Brown’s early groundbreaking findings in neuroscience included a novel algorithm that decodes the position of an animal by observing the activity of a small group of place cells in the animal’s brain, a discovery he made while working with fellow Picower Institute investigator Matt Wilson in the 1990s. The resulting state-space algorithm for point processes not only offered much better decoding with fewer neurons than previous approaches, but it also established a new framework for specifying dynamically the relationship between the spike trains (the timing sequence of firing neurons) in the brain and factors from the outside world.

      “One of the basic questions at the time was whether an animal holds a representation of where it is in its mind — in the hippocampus,” Brown says. “We were able to show that it did, and we could show that with only 30 neurons.”

      After introducing this state-space paradigm to neuroscience, Brown went on to refine the original idea and apply it to other dynamic situations — to simultaneously track neural activity and learning, for example, and to define with precision anesthesia-induced loss of consciousness, as well as its subsequent recovery. In the early 2000s, Brown put together a team to specifically study anesthesia’s effects on the brain.

      Through experimental research and mathematical modeling, Brown and his team showed that the altered arousal states produced by the main classes of anesthesia medications can be characterized by analyzing the oscillatory patterns observed in the EEG along with the locations of their molecular targets, and the anatomy and physiology of the neural circuits that connect those locations. He has established, including in recent papers with Picower Professor Earl K. Miller, that a principal way in which anesthetics produce unconsciousness is by producing oscillations that impair how different brain regions communicate with each other.

      The result of Brown’s research has been a new paradigm for brain monitoring during general anesthesia for surgery, one that allows an anesthesiologist to dose the patient based on EEG readouts (neural oscillations) of the patient’s anesthetic state rather than purely on vital sign responses. This pioneering approach promises to revolutionize how anesthesia medications are delivered to patients, and also shed light on other altered states of arousal such as sleep and coma.

      To advance that vision, Brown recently discussed how he is working to develop a new research center at MIT and MGH to further integrate anesthesiology with neuroscience research. The Brain Arousal State Control Innovation Center, he said, would not only advance anesthesiology care but also harness insights gained from anesthesiology research to improve other aspects of clinical neuroscience.

      “By demonstrating that physics and mathematics can make an enormous contribution to neuroscience, doctors Abbott, Brown, Sejnowski, and Sompolinsky have inspired an entire new generation of physicists and other quantitative scientists to follow in their footsteps,” says Frances Jensen, professor and chair of the Department of Neurology and co-director of the Penn Medicine Translational Neuroscience Center within the Perelman School of Medicine at the University of Pennsylvania, and chair of the Selection Advisory Board to the prize. “The ramifications for neuroscience have been broad and profound. It is a great pleasure to be honoring each of them with this prestigious award.”

      This report was adapted from materials provided by the Gruber Foundation. More

    • 113 Shares169 Views

      in Data Management & Statistics

      Artificial intelligence predicts patients’ race from their medical images

      by

      The miseducation of algorithms is a critical problem; when artificial intelligence mirrors unconscious thoughts, racism, and biases of the humans who generated these algorithms, it can lead to serious harm. Computer programs, for example, have wrongly flagged Black defendants as twice as likely to reoffend as someone who’s white. When an AI used cost as a proxy for health needs, it falsely named Black patients as healthier than equally sick white ones, as less money was spent on them. Even AI used to write a play relied on using harmful stereotypes for casting. 

      Removing sensitive features from the data seems like a viable tweak. But what happens when it’s not enough? 

      Examples of bias in natural language processing are boundless — but MIT scientists have investigated another important, largely underexplored modality: medical images. Using both private and public datasets, the team found that AI can accurately predict self-reported race of patients from medical images alone. Using imaging data of chest X-rays, limb X-rays, chest CT scans, and mammograms, the team trained a deep learning model to identify race as white, Black, or Asian — even though the images themselves contained no explicit mention of the patient’s race. This is a feat even the most seasoned physicians cannot do, and it’s not clear how the model was able to do this. 

      In an attempt to tease out and make sense of the enigmatic “how” of it all, the researchers ran a slew of experiments. To investigate possible mechanisms of race detection, they looked at variables like differences in anatomy, bone density, resolution of images — and many more, and the models still prevailed with high ability to detect race from chest X-rays. “These results were initially confusing, because the members of our research team could not come anywhere close to identifying a good proxy for this task,” says paper co-author Marzyeh Ghassemi, an assistant professor in the MIT Department of Electrical Engineering and Computer Science and the Institute for Medical Engineering and Science (IMES), who is an affiliate of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and of the MIT Jameel Clinic. “Even when you filter medical images past where the images are recognizable as medical images at all, deep models maintain a very high performance. That is concerning because superhuman capacities are generally much more difficult to control, regulate, and prevent from harming people.”

      In a clinical setting, algorithms can help tell us whether a patient is a candidate for chemotherapy, dictate the triage of patients, or decide if a movement to the ICU is necessary. “We think that the algorithms are only looking at vital signs or laboratory tests, but it’s possible they’re also looking at your race, ethnicity, sex, whether you’re incarcerated or not — even if all of that information is hidden,” says paper co-author Leo Anthony Celi, principal research scientist in IMES at MIT and associate professor of medicine at Harvard Medical School. “Just because you have representation of different groups in your algorithms, that doesn’t guarantee it won’t perpetuate or magnify existing disparities and inequities. Feeding the algorithms with more data with representation is not a panacea. This paper should make us pause and truly reconsider whether we are ready to bring AI to the bedside.” 

      The study, “AI recognition of patient race in medical imaging: a modeling study,” was published in Lancet Digital Health on May 11. Celi and Ghassemi wrote the paper alongside 20 other authors in four countries.

      To set up the tests, the scientists first showed that the models were able to predict race across multiple imaging modalities, various datasets, and diverse clinical tasks, as well as across a range of academic centers and patient populations in the United States. They used three large chest X-ray datasets, and tested the model on an unseen subset of the dataset used to train the model and a completely different one. Next, they trained the racial identity detection models for non-chest X-ray images from multiple body locations, including digital radiography, mammography, lateral cervical spine radiographs, and chest CTs to see whether the model’s performance was limited to chest X-rays. 

      The team covered many bases in an attempt to explain the model’s behavior: differences in physical characteristics between different racial groups (body habitus, breast density), disease distribution (previous studies have shown that Black patients have a higher incidence for health issues like cardiac disease), location-specific or tissue specific differences, effects of societal bias and environmental stress, the ability of deep learning systems to detect race when multiple demographic and patient factors were combined, and if specific image regions contributed to recognizing race. 

      What emerged was truly staggering: The ability of the models to predict race from diagnostic labels alone was much lower than the chest X-ray image-based models. 

      For example, the bone density test used images where the thicker part of the bone appeared white, and the thinner part appeared more gray or translucent. Scientists assumed that since Black people generally have higher bone mineral density, the color differences helped the AI models to detect race. To cut that off, they clipped the images with a filter, so the model couldn’t color differences. It turned out that cutting off the color supply didn’t faze the model — it still could accurately predict races. (The “Area Under the Curve” value, meaning the measure of the accuracy of a quantitative diagnostic test, was 0.94–0.96). As such, the learned features of the model appeared to rely on all regions of the image, meaning that controlling this type of algorithmic behavior presents a messy, challenging problem. 

      The scientists acknowledge limited availability of racial identity labels, which caused them to focus on Asian, Black, and white populations, and that their ground truth was a self-reported detail. Other forthcoming work will include potentially looking at isolating different signals before image reconstruction, because, as with bone density experiments, they couldn’t account for residual bone tissue that was on the images. 

      Notably, other work by Ghassemi and Celi led by MIT student Hammaad Adam has found that models can also identify patient self-reported race from clinical notes even when those notes are stripped of explicit indicators of race. Just as in this work, human experts are not able to accurately predict patient race from the same redacted clinical notes.

      “We need to bring social scientists into the picture. Domain experts, which are usually the clinicians, public health practitioners, computer scientists, and engineers are not enough. Health care is a social-cultural problem just as much as it’s a medical problem. We need another group of experts to weigh in and to provide input and feedback on how we design, develop, deploy, and evaluate these algorithms,” says Celi. “We need to also ask the data scientists, before any exploration of the data, are there disparities? Which patient groups are marginalized? What are the drivers of those disparities? Is it access to care? Is it from the subjectivity of the care providers? If we don’t understand that, we won’t have a chance of being able to identify the unintended consequences of the algorithms, and there’s no way we’ll be able to safeguard the algorithms from perpetuating biases.”

      “The fact that algorithms ‘see’ race, as the authors convincingly document, can be dangerous. But an important and related fact is that, when used carefully, algorithms can also work to counter bias,” says Ziad Obermeyer, associate professor at the University of California at Berkeley, whose research focuses on AI applied to health. “In our own work, led by computer scientist Emma Pierson at Cornell, we show that algorithms that learn from patients’ pain experiences can find new sources of knee pain in X-rays that disproportionately affect Black patients — and are disproportionately missed by radiologists. So just like any tool, algorithms can be a force for evil or a force for good — which one depends on us, and the choices we make when we build algorithms.”

      The work is supported, in part, by the National Institutes of Health. More

    • 50 Shares109 Views

      in Data Management & Statistics

      Living better with algorithms

      by

      Laboratory for Information and Decision Systems (LIDS) student Sarah Cen remembers the lecture that sent her down the track to an upstream question.

      At a talk on ethical artificial intelligence, the speaker brought up a variation on the famous trolley problem, which outlines a philosophical choice between two undesirable outcomes.

      The speaker’s scenario: Say a self-driving car is traveling down a narrow alley with an elderly woman walking on one side and a small child on the other, and no way to thread between both without a fatality. Who should the car hit?

      Then the speaker said: Let’s take a step back. Is this the question we should even be asking?

      That’s when things clicked for Cen. Instead of considering the point of impact, a self-driving car could have avoided choosing between two bad outcomes by making a decision earlier on — the speaker pointed out that, when entering the alley, the car could have determined that the space was narrow and slowed to a speed that would keep everyone safe.

      Recognizing that today’s AI safety approaches often resemble the trolley problem, focusing on downstream regulation such as liability after someone is left with no good choices, Cen wondered: What if we could design better upstream and downstream safeguards to such problems? This question has informed much of Cen’s work.

      “Engineering systems are not divorced from the social systems on which they intervene,” Cen says. Ignoring this fact risks creating tools that fail to be useful when deployed or, more worryingly, that are harmful.

      Cen arrived at LIDS in 2018 via a slightly roundabout route. She first got a taste for research during her undergraduate degree at Princeton University, where she majored in mechanical engineering. For her master’s degree, she changed course, working on radar solutions in mobile robotics (primarily for self-driving cars) at Oxford University. There, she developed an interest in AI algorithms, curious about when and why they misbehave. So, she came to MIT and LIDS for her doctoral research, working with Professor Devavrat Shah in the Department of Electrical Engineering and Computer Science, for a stronger theoretical grounding in information systems.

      Auditing social media algorithms

      Together with Shah and other collaborators, Cen has worked on a wide range of projects during her time at LIDS, many of which tie directly to her interest in the interactions between humans and computational systems. In one such project, Cen studies options for regulating social media. Her recent work provides a method for translating human-readable regulations into implementable audits.

      To get a sense of what this means, suppose that regulators require that any public health content — for example, on vaccines — not be vastly different for politically left- and right-leaning users. How should auditors check that a social media platform complies with this regulation? Can a platform be made to comply with the regulation without damaging its bottom line? And how does compliance affect the actual content that users do see?

      Designing an auditing procedure is difficult in large part because there are so many stakeholders when it comes to social media. Auditors have to inspect the algorithm without accessing sensitive user data. They also have to work around tricky trade secrets, which can prevent them from getting a close look at the very algorithm that they are auditing because these algorithms are legally protected. Other considerations come into play as well, such as balancing the removal of misinformation with the protection of free speech.

      To meet these challenges, Cen and Shah developed an auditing procedure that does not need more than black-box access to the social media algorithm (which respects trade secrets), does not remove content (which avoids issues of censorship), and does not require access to users (which preserves users’ privacy).

      In their design process, the team also analyzed the properties of their auditing procedure, finding that it ensures a desirable property they call decision robustness. As good news for the platform, they show that a platform can pass the audit without sacrificing profits. Interestingly, they also found the audit naturally incentivizes the platform to show users diverse content, which is known to help reduce the spread of misinformation, counteract echo chambers, and more.

      Who gets good outcomes and who gets bad ones?

      In another line of research, Cen looks at whether people can receive good long-term outcomes when they not only compete for resources, but also don’t know upfront what resources are best for them.

      Some platforms, such as job-search platforms or ride-sharing apps, are part of what is called a matching market, which uses an algorithm to match one set of individuals (such as workers or riders) with another (such as employers or drivers). In many cases, individuals have matching preferences that they learn through trial and error. In labor markets, for example, workers learn their preferences about what kinds of jobs they want, and employers learn their preferences about the qualifications they seek from workers.

      But learning can be disrupted by competition. If workers with a particular background are repeatedly denied jobs in tech because of high competition for tech jobs, for instance, they may never get the knowledge they need to make an informed decision about whether they want to work in tech. Similarly, tech employers may never see and learn what these workers could do if they were hired.

      Cen’s work examines this interaction between learning and competition, studying whether it is possible for individuals on both sides of the matching market to walk away happy.

      Modeling such matching markets, Cen and Shah found that it is indeed possible to get to a stable outcome (workers aren’t incentivized to leave the matching market), with low regret (workers are happy with their long-term outcomes), fairness (happiness is evenly distributed), and high social welfare.

      Interestingly, it’s not obvious that it’s possible to get stability, low regret, fairness, and high social welfare simultaneously.  So another important aspect of the research was uncovering when it is possible to achieve all four criteria at once and exploring the implications of those conditions.

      What is the effect of X on Y?

      For the next few years, though, Cen plans to work on a new project, studying how to quantify the effect of an action X on an outcome Y when it’s expensive — or impossible — to measure this effect, focusing in particular on systems that have complex social behaviors.

      For instance, when Covid-19 cases surged in the pandemic, many cities had to decide what restrictions to adopt, such as mask mandates, business closures, or stay-home orders. They had to act fast and balance public health with community and business needs, public spending, and a host of other considerations.

      Typically, in order to estimate the effect of restrictions on the rate of infection, one might compare the rates of infection in areas that underwent different interventions. If one county has a mask mandate while its neighboring county does not, one might think comparing the counties’ infection rates would reveal the effectiveness of mask mandates. 

      But of course, no county exists in a vacuum. If, for instance, people from both counties gather to watch a football game in the maskless county every week, people from both counties mix. These complex interactions matter, and Sarah plans to study questions of cause and effect in such settings.

      “We’re interested in how decisions or interventions affect an outcome of interest, such as how criminal justice reform affects incarceration rates or how an ad campaign might change the public’s behaviors,” Cen says.

      Cen has also applied the principles of promoting inclusivity to her work in the MIT community.

      As one of three co-presidents of the Graduate Women in MIT EECS student group, she helped organize the inaugural GW6 research summit featuring the research of women graduate students — not only to showcase positive role models to students, but also to highlight the many successful graduate women at MIT who are not to be underestimated.

      Whether in computing or in the community, a system taking steps to address bias is one that enjoys legitimacy and trust, Cen says. “Accountability, legitimacy, trust — these principles play crucial roles in society and, ultimately, will determine which systems endure with time.”  More

    • 88 Shares99 Views

      in Data Management & Statistics

      Technique protects privacy when making online recommendations

      by

      Algorithms recommend products while we shop online or suggest songs we might like as we listen to music on streaming apps.

      These algorithms work by using personal information like our past purchases and browsing history to generate tailored recommendations. The sensitive nature of such data makes preserving privacy extremely important, but existing methods for solving this problem rely on heavy cryptographic tools requiring enormous amounts of computation and bandwidth.

      MIT researchers may have a better solution. They developed a privacy-preserving protocol that is so efficient it can run on a smartphone over a very slow network. Their technique safeguards personal data while ensuring recommendation results are accurate.

      In addition to user privacy, their protocol minimizes the unauthorized transfer of information from the database, known as leakage, even if a malicious agent tries to trick a database into revealing secret information.

      The new protocol could be especially useful in situations where data leaks could violate user privacy laws, like when a health care provider uses a patient’s medical history to search a database for other patients who had similar symptoms or when a company serves targeted advertisements to users under European privacy regulations.

      “This is a really hard problem. We relied on a whole string of cryptographic and algorithmic tricks to arrive at our protocol,” says Sacha Servan-Schreiber, a graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and lead author of the paper that presents this new protocol.

      Servan-Schreiber wrote the paper with fellow CSAIL graduate student Simon Langowski and their advisor and senior author Srinivas Devadas, the Edwin Sibley Webster Professor of Electrical Engineering. The research will be presented at the IEEE Symposium on Security and Privacy.

      The data next door

      The technique at the heart of algorithmic recommendation engines is known as a nearest neighbor search, which involves finding the data point in a database that is closest to a query point. Data points that are mapped nearby share similar attributes and are called neighbors.

      These searches involve a server that is linked with an online database which contains concise representations of data point attributes. In the case of a music streaming service, those attributes, known as feature vectors, could be the genre or popularity of different songs.

      To find a song recommendation, the client (user) sends a query to the server that contains a certain feature vector, like a genre of music the user likes or a compressed history of their listening habits. The server then provides the ID of a feature vector in the database that is closest to the client’s query, without revealing the actual vector. In the case of music streaming, that ID would likely be a song title. The client learns the recommended song title without learning the feature vector associated with it.

      “The server has to be able to do this computation without seeing the numbers it is doing the computation on. It can’t actually see the features, but still needs to give you the closest thing in the database,” says Langowski.

      To achieve this, the researchers created a protocol that relies on two separate servers that access the same database. Using two servers makes the process more efficient and enables the use of a cryptographic technique known as private information retrieval. This technique allows a client to query a database without revealing what it is searching for, Servan-Schreiber explains.

      Overcoming security challenges

      But while private information retrieval is secure on the client side, it doesn’t provide database privacy on its own. The database offers a set of candidate vectors — possible nearest neighbors — for the client, which are typically winnowed down later by the client using brute force. However, doing so can reveal a lot about the database to the client. The additional privacy challenge is to prevent the client from learning those extra vectors. 

      The researchers employed a tuning technique that eliminates many of the extra vectors in the first place, and then used a different trick, which they call oblivious masking, to hide any additional data points except for the actual nearest neighbor. This efficiently preserves database privacy, so the client won’t learn anything about the feature vectors in the database.  

      Once they designed this protocol, they tested it with a nonprivate implementation on four real-world datasets to determine how to tune the algorithm to maximize accuracy. Then, they used their protocol to conduct private nearest neighbor search queries on those datasets.

      Their technique requires a few seconds of server processing time per query and less than 10 megabytes of communication between the client and servers, even with databases that contained more than 10 million items. By contrast, other secure methods can require gigabytes of communication or hours of computation time. With each query, their method achieved greater than 95 percent accuracy (meaning that nearly every time it found the actual approximate nearest neighbor to the query point). 

      The techniques they used to enable database privacy will thwart a malicious client even if it sends false queries to try and trick the server into leaking information.

      “A malicious client won’t learn much more information than an honest client following protocol. And it protects against malicious servers, too. If one deviates from protocol, you might not get the right result, but they will never learn what the client’s query was,” Langowski says.

      In the future, the researchers plan to adjust the protocol so it can preserve privacy using only one server. This could enable it to be applied in more real-world situations, since it would not require the use of two noncolluding entities (which don’t share information with each other) to manage the database.  

      “Nearest neighbor search undergirds many critical machine-learning driven applications, from providing users with content recommendations to classifying medical conditions. However, it typically requires sharing a lot of data with a central system to aggregate and enable the search,” says Bayan Bruss, head of applied machine-learning research at Capital One, who was not involved with this work. “This research provides a key step towards ensuring that the user receives the benefits from nearest neighbor search while having confidence that the central system will not use their data for other purposes.” More

    • 63 Shares99 Views

      in Data Management & Statistics

      Estimating the informativeness of data

      by

      Not all data are created equal. But how much information is any piece of data likely to contain? This question is central to medical testing, designing scientific experiments, and even to everyday human learning and thinking. MIT researchers have developed a new way to solve this problem, opening up new applications in medicine, scientific discovery, cognitive science, and artificial intelligence.

      In theory, the 1948 paper, “A Mathematical Theory of Communication,” by the late MIT Professor Emeritus Claude Shannon answered this question definitively. One of Shannon’s breakthrough results is the idea of entropy, which lets us quantify the amount of information inherent in any random object, including random variables that model observed data. Shannon’s results created the foundations of information theory and modern telecommunications. The concept of entropy has also proven central to computer science and machine learning.

      The challenge of estimating entropy

      Unfortunately, the use of Shannon’s formula can quickly become computationally intractable. It requires precisely calculating the probability of the data, which in turn requires calculating every possible way the data could have arisen under a probabilistic model. If the data-generating process is very simple — for example, a single toss of a coin or roll of a loaded die — then calculating entropies is straightforward. But consider the problem of medical testing, where a positive test result is the result of hundreds of interacting variables, all unknown. With just 10 unknowns, there are already 1,000 possible explanations for the data. With a few hundred, there are more possible explanations than atoms in the known universe, which makes calculating the entropy exactly an unmanageable problem.

      MIT researchers have developed a new method to estimate good approximations to many information quantities such as Shannon entropy by using probabilistic inference. The work appears in a paper presented at AISTATS 2022 by authors Feras Saad ’16, MEng ’16, a PhD candidate in electrical engineering and computer science; Marco-Cusumano Towner PhD ’21; and Vikash Mansinghka ’05, MEng ’09, PhD ’09, a principal research scientist in the Department of Brain and Cognitive Sciences. The key insight is, rather than enumerate all explanations, to instead use probabilistic inference algorithms to first infer which explanations are probable and then use these probable explanations to construct high-quality entropy estimates. The paper shows that this inference-based approach can be much faster and more accurate than previous approaches.

      Estimating entropy and information in a probabilistic model is fundamentally hard because it often requires solving a high-dimensional integration problem. Many previous works have developed estimators of these quantities for certain special cases, but the new estimators of entropy via inference (EEVI) offer the first approach that can deliver sharp upper and lower bounds on a broad set of information-theoretic quantities. An upper and lower bound means that although we don’t know the true entropy, we can get a number that is smaller than it and a number that is higher than it.

      “The upper and lower bounds on entropy delivered by our method are particularly useful for three reasons,” says Saad. “First, the difference between the upper and lower bounds gives a quantitative sense of how confident we should be about the estimates. Second, by using more computational effort we can drive the difference between the two bounds to zero, which ‘squeezes’ the true value with a high degree of accuracy. Third, we can compose these bounds to form estimates of many other quantities that tell us how informative different variables in a model are of one another.”

      Solving fundamental problems with data-driven expert systems

      Saad says he is most excited about the possibility that this method gives for querying probabilistic models in areas like machine-assisted medical diagnoses. He says one goal of the EEVI method is to be able to solve new queries using rich generative models for things like liver disease and diabetes that have already been developed by experts in the medical domain. For example, suppose we have a patient with a set of observed attributes (height, weight, age, etc.) and observed symptoms (nausea, blood pressure, etc.). Given these attributes and symptoms, EEVI can be used to help determine which medical tests for symptoms the physician should conduct to maximize information about the absence or presence of a given liver disease (like cirrhosis or primary biliary cholangitis).

      For insulin diagnosis, the authors showed how to use the method for computing optimal times to take blood glucose measurements that maximize information about a patient’s insulin sensitivity, given an expert-built probabilistic model of insulin metabolism and the patient’s personalized meal and medication schedule. As routine medical tracking like glucose monitoring moves away from doctor’s offices and toward wearable devices, there are even more opportunities to improve data acquisition, if the value of the data can be estimated accurately in advance.

      Vikash Mansinghka, senior author on the paper, adds, “We’ve shown that probabilistic inference algorithms can be used to estimate rigorous bounds on information measures that AI engineers often think of as intractable to calculate. This opens up many new applications. It also shows that inference may be more computationally fundamental than we thought. It also helps to explain how human minds might be able to estimate the value of information so pervasively, as a central building block of everyday cognition, and help us engineer AI expert systems that have these capabilities.”

      The paper, “Estimators of Entropy and Information via Inference in Probabilistic Models,” was presented at AISTATS 2022. More

    • 113 Shares179 Views

      in Data Management & Statistics

      A new state of the art for unsupervised vision

      by

      Labeling data can be a chore. It’s the main source of sustenance for computer-vision models; without it, they’d have a lot of difficulty identifying objects, people, and other important image characteristics. Yet producing just an hour of tagged and labeled data can take a whopping 800 hours of human time. Our high-fidelity understanding of the world develops as machines can better perceive and interact with our surroundings. But they need more help.

      Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), Microsoft, and Cornell University have attempted to solve this problem plaguing vision models by creating “STEGO,” an algorithm that can jointly discover and segment objects without any human labels at all, down to the pixel.

      STEGO learns something called “semantic segmentation” — fancy speak for the process of assigning a label to every pixel in an image. Semantic segmentation is an important skill for today’s computer-vision systems because images can be cluttered with objects. Even more challenging is that these objects don’t always fit into literal boxes; algorithms tend to work better for discrete “things” like people and cars as opposed to “stuff” like vegetation, sky, and mashed potatoes. A previous system might simply perceive a nuanced scene of a dog playing in the park as just a dog, but by assigning every pixel of the image a label, STEGO can break the image into its main ingredients: a dog, sky, grass, and its owner.

      Play video

      A new state of the art for unsupervised computer vision

      Assigning every single pixel of the world a label is ambitious — especially without any kind of feedback from humans. The majority of algorithms today get their knowledge from mounds of labeled data, which can take painstaking human-hours to source. Just imagine the excitement of labeling every pixel of 100,000 images! To discover these objects without a human’s helpful guidance, STEGO looks for similar objects that appear throughout a dataset. It then associates these similar objects together to construct a consistent view of the world across all of the images it learns from.

      Seeing the world

      Machines that can “see” are crucial for a wide array of new and emerging technologies like self-driving cars and predictive modeling for medical diagnostics. Since STEGO can learn without labels, it can detect objects in many different domains, even those that humans don’t yet understand fully. 

      “If you’re looking at oncological scans, the surface of planets, or high-resolution biological images, it’s hard to know what objects to look for without expert knowledge. In emerging domains, sometimes even human experts don’t know what the right objects should be,” says Mark Hamilton, a PhD student in electrical engineering and computer science at MIT, research affiliate of MIT CSAIL, software engineer at Microsoft, and lead author on a new paper about STEGO. “In these types of situations where you want to design a method to operate at the boundaries of science, you can’t rely on humans to figure it out before machines do.”

      STEGO was tested on a slew of visual domains spanning general images, driving images, and high-altitude aerial photographs. In each domain, STEGO was able to identify and segment relevant objects that were closely aligned with human judgments. STEGO’s most diverse benchmark was the COCO-Stuff dataset, which is made up of diverse images from all over the world, from indoor scenes to people playing sports to trees and cows. In most cases, the previous state-of-the-art system could capture a low-resolution gist of a scene, but struggled on fine-grained details: A human was a blob, a motorcycle was captured as a person, and it couldn’t recognize any geese. On the same scenes, STEGO doubled the performance of previous systems and discovered concepts like animals, buildings, people, furniture, and many others.

      STEGO not only doubled the performance of prior systems on the COCO-Stuff benchmark, but made similar leaps forward in other visual domains. When applied to driverless car datasets, STEGO successfully segmented out roads, people, and street signs with much higher resolution and granularity than previous systems. On images from space, the system broke down every single square foot of the surface of the Earth into roads, vegetation, and buildings. 

      Connecting the pixels

      STEGO — which stands for “Self-supervised Transformer with Energy-based Graph Optimization” — builds on top of the DINO algorithm, which learned about the world through 14 million images from the ImageNet database. STEGO refines the DINO backbone through a learning process that mimics our own way of stitching together pieces of the world to make meaning. 

      For example, you might consider two images of dogs walking in the park. Even though they’re different dogs, with different owners, in different parks, STEGO can tell (without humans) how each scene’s objects relate to each other. The authors even probe STEGO’s mind to see how each little, brown, furry thing in the images are similar, and likewise with other shared objects like grass and people. By connecting objects across images, STEGO builds a consistent view of the word.

      “The idea is that these types of algorithms can find consistent groupings in a largely automated fashion so we don’t have to do that ourselves,” says Hamilton. “It might have taken years to understand complex visual datasets like biological imagery, but if we can avoid spending 1,000 hours combing through data and labeling it, we can find and discover new information that we might have missed. We hope this will help us understand the visual word in a more empirically grounded way.”

      Looking ahead

      Despite its improvements, STEGO still faces certain challenges. One is that labels can be arbitrary. For example, the labels of the COCO-Stuff dataset distinguish between “food-things” like bananas and chicken wings, and “food-stuff” like grits and pasta. STEGO doesn’t see much of a distinction there. In other cases, STEGO was confused by odd images — like one of a banana sitting on a phone receiver — where the receiver was labeled “foodstuff,” instead of “raw material.” 

      For upcoming work, they’re planning to explore giving STEGO a bit more flexibility than just labeling pixels into a fixed number of classes as things in the real world can sometimes be multiple things at the same time (like “food”, “plant” and “fruit”). The authors hope this will give the algorithm room for uncertainty, trade-offs, and more abstract thinking.

      “In making a general tool for understanding potentially complicated datasets, we hope that this type of an algorithm can automate the scientific process of object discovery from images. There’s a lot of different domains where human labeling would be prohibitively expensive, or humans simply don’t even know the specific structure, like in certain biological and astrophysical domains. We hope that future work enables application to a very broad scope of datasets. Since you don’t need any human labels, we can now start to apply ML tools more broadly,” says Hamilton.

      “STEGO is simple, elegant, and very effective. I consider unsupervised segmentation to be a benchmark for progress in image understanding, and a very difficult problem. The research community has made terrific progress in unsupervised image understanding with the adoption of transformer architectures,” says Andrea Vedaldi, professor of computer vision and machine learning and a co-lead of the Visual Geometry Group at the engineering science department of the University of Oxford. “This research provides perhaps the most direct and effective demonstration of this progress on unsupervised segmentation.” 

      Hamilton wrote the paper alongside MIT CSAIL PhD student Zhoutong Zhang, Assistant Professor Bharath Hariharan of Cornell University, Associate Professor Noah Snavely of Cornell Tech, and MIT professor William T. Freeman. They will present the paper at the 2022 International Conference on Learning Representations (ICLR).  More