Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 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

    • 138 Shares189 Views

      in Data Management & Statistics

      Artificial intelligence system learns concepts shared across video, audio, and text

      by

      Humans observe the world through a combination of different modalities, like vision, hearing, and our understanding of language. Machines, on the other hand, interpret the world through data that algorithms can process.

      So, when a machine “sees” a photo, it must encode that photo into data it can use to perform a task like image classification. This process becomes more complicated when inputs come in multiple formats, like videos, audio clips, and images.

      “The main challenge here is, how can a machine align those different modalities? As humans, this is easy for us. We see a car and then hear the sound of a car driving by, and we know these are the same thing. But for machine learning, it is not that straightforward,” says Alexander Liu, a graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and first author of a paper tackling this problem. 

      Liu and his collaborators developed an artificial intelligence technique that learns to represent data in a way that captures concepts which are shared between visual and audio modalities. For instance, their method can learn that the action of a baby crying in a video is related to the spoken word “crying” in an audio clip.

      Using this knowledge, their machine-learning model can identify where a certain action is taking place in a video and label it.

      It performs better than other machine-learning methods at cross-modal retrieval tasks, which involve finding a piece of data, like a video, that matches a user’s query given in another form, like spoken language. Their model also makes it easier for users to see why the machine thinks the video it retrieved matches their query.

      This technique could someday be utilized to help robots learn about concepts in the world through perception, more like the way humans do.

      Joining Liu on the paper are CSAIL postdoc SouYoung Jin; grad students Cheng-I Jeff Lai and Andrew Rouditchenko; Aude Oliva, senior research scientist in CSAIL and MIT director of the MIT-IBM Watson AI Lab; and senior author James Glass, senior research scientist and head of the Spoken Language Systems Group in CSAIL. The research will be presented at the Annual Meeting of the Association for Computational Linguistics.

      Learning representations

      The researchers focus their work on representation learning, which is a form of machine learning that seeks to transform input data to make it easier to perform a task like classification or prediction.

      The representation learning model takes raw data, such as videos and their corresponding text captions, and encodes them by extracting features, or observations about objects and actions in the video. Then it maps those data points in a grid, known as an embedding space. The model clusters similar data together as single points in the grid. Each of these data points, or vectors, is represented by an individual word.

      For instance, a video clip of a person juggling might be mapped to a vector labeled “juggling.”

      The researchers constrain the model so it can only use 1,000 words to label vectors. The model can decide which actions or concepts it wants to encode into a single vector, but it can only use 1,000 vectors. The model chooses the words it thinks best represent the data.

      Rather than encoding data from different modalities onto separate grids, their method employs a shared embedding space where two modalities can be encoded together. This enables the model to learn the relationship between representations from two modalities, like video that shows a person juggling and an audio recording of someone saying “juggling.”

      To help the system process data from multiple modalities, they designed an algorithm that guides the machine to encode similar concepts into the same vector.

      “If there is a video about pigs, the model might assign the word ‘pig’ to one of the 1,000 vectors. Then if the model hears someone saying the word ‘pig’ in an audio clip, it should still use the same vector to encode that,” Liu explains.

      A better retriever

      They tested the model on cross-modal retrieval tasks using three datasets: a video-text dataset with video clips and text captions, a video-audio dataset with video clips and spoken audio captions, and an image-audio dataset with images and spoken audio captions.

      For example, in the video-audio dataset, the model chose 1,000 words to represent the actions in the videos. Then, when the researchers fed it audio queries, the model tried to find the clip that best matched those spoken words.

      “Just like a Google search, you type in some text and the machine tries to tell you the most relevant things you are searching for. Only we do this in the vector space,” Liu says.

      Not only was their technique more likely to find better matches than the models they compared it to, it is also easier to understand.

      Because the model could only use 1,000 total words to label vectors, a user can more see easily which words the machine used to conclude that the video and spoken words are similar. This could make the model easier to apply in real-world situations where it is vital that users understand how it makes decisions, Liu says.

      The model still has some limitations they hope to address in future work. For one, their research focused on data from two modalities at a time, but in the real world humans encounter many data modalities simultaneously, Liu says.

      “And we know 1,000 words works on this kind of dataset, but we don’t know if it can be generalized to a real-world problem,” he adds.

      Plus, the images and videos in their datasets contained simple objects or straightforward actions; real-world data are much messier. They also want to determine how well their method scales up when there is a wider diversity of inputs.

      This research was supported, in part, by the MIT-IBM Watson AI Lab and its member companies, Nexplore and Woodside, and by the MIT Lincoln Laboratory. More

    • 50 Shares189 Views

      in Data Management & Statistics

      What words can convey

      by

      From search engines to voice assistants, computers are getting better at understanding what we mean. That’s thanks to language-processing programs that make sense of a staggering number of words, without ever being told explicitly what those words mean. Such programs infer meaning instead through statistics — and a new study reveals that this computational approach can assign many kinds of information to a single word, just like the human brain.

      The study, published April 14 in the journal Nature Human Behavior, was co-led by Gabriel Grand, a graduate student in electrical engineering and computer science who is affiliated with MIT’s Computer Science and Artificial Intelligence Laboratory, and Idan Blank PhD ’16, an assistant professor at the University of California at Los Angeles. The work was supervised by McGovern Institute for Brain Research investigator Ev Fedorenko, a cognitive neuroscientist who studies how the human brain uses and understands language, and Francisco Pereira at the National Institute of Mental Health. Fedorenko says the rich knowledge her team was able to find within computational language models demonstrates just how much can be learned about the world through language alone.

      The research team began its analysis of statistics-based language processing models in 2015, when the approach was new. Such models derive meaning by analyzing how often pairs of words co-occur in texts and using those relationships to assess the similarities of words’ meanings. For example, such a program might conclude that “bread” and “apple” are more similar to one another than they are to “notebook,” because “bread” and “apple” are often found in proximity to words like “eat” or “snack,” whereas “notebook” is not.

      The models were clearly good at measuring words’ overall similarity to one another. But most words carry many kinds of information, and their similarities depend on which qualities are being evaluated. “Humans can come up with all these different mental scales to help organize their understanding of words,” explains Grand, a former undergraduate researcher in the Fedorenko lab. For example, he says, “dolphins and alligators might be similar in size, but one is much more dangerous than the other.”

      Grand and Blank, who was then a graduate student at the McGovern Institute, wanted to know whether the models captured that same nuance. And if they did, how was the information organized?

      To learn how the information in such a model stacked up to humans’ understanding of words, the team first asked human volunteers to score words along many different scales: Were the concepts those words conveyed big or small, safe or dangerous, wet or dry? Then, having mapped where people position different words along these scales, they looked to see whether language processing models did the same.

      Grand explains that distributional semantic models use co-occurrence statistics to organize words into a huge, multidimensional matrix. The more similar words are to one another, the closer they are within that space. The dimensions of the space are vast, and there is no inherent meaning built into its structure. “In these word embeddings, there are hundreds of dimensions, and we have no idea what any dimension means,” he says. “We’re really trying to peer into this black box and say, ‘is there structure in here?’”

      Specifically, they asked whether the semantic scales they had asked their volunteers use were represented in the model. So they looked to see where words in the space lined up along vectors defined by the extremes of those scales. Where did dolphins and tigers fall on line from “big” to “small,” for example? And were they closer together along that line than they were on a line representing danger (“safe” to “dangerous”)?

      Across more than 50 sets of world categories and semantic scales, they found that the model had organized words very much like the human volunteers. Dolphins and tigers were judged to be similar in terms of size, but far apart on scales measuring danger or wetness. The model had organized the words in a way that represented many kinds of meaning — and it had done so based entirely on the words’ co-occurrences.

      That, Fedorenko says, tells us something about the power of language. “The fact that we can recover so much of this rich semantic information from just these simple word co-occurrence statistics suggests that this is one very powerful source of learning about things that you may not even have direct perceptual experience with.” More

    • 163 Shares119 Views

      in Data Management & Statistics

      Seven from MIT elected to American Academy of Arts and Sciences for 2022

      by

      Seven MIT faculty members are among more than 250 leaders from academia, the arts, industry, public policy, and research elected to the American Academy of Arts and Sciences, the academy announced Thursday.

      One of the nation’s most prestigious honorary societies, the academy is also a leading center for independent policy research. Members contribute to academy publications, as well as studies of science and technology policy, energy and global security, social policy and American institutions, the humanities and culture, and education.

      Those elected from MIT this year are:

      Alberto Abadie, professor of economics and associate director of the Institute for Data, Systems, and Society
      Regina Barzilay, the School of Engineering Distinguished Professor for AI and Health
      Roman Bezrukavnikov, professor of mathematics
      Michale S. Fee, the Glen V. and Phyllis F. Dorflinger Professor and head of the Department of Brain and Cognitive Sciences
      Dina Katabi, the Thuan and Nicole Pham Professor
      Ronald T. Raines, the Roger and Georges Firmenich Professor of Natural Products Chemistry
      Rebecca R. Saxe, the John W. Jarve Professor of Brain and Cognitive Sciences

      “We are celebrating a depth of achievements in a breadth of areas,” says David Oxtoby, president of the American Academy. “These individuals excel in ways that excite us and inspire us at a time when recognizing excellence, commending expertise, and working toward the common good is absolutely essential to realizing a better future.”

      Since its founding in 1780, the academy has elected leading thinkers from each generation, including George Washington and Benjamin Franklin in the 18th century, Maria Mitchell and Daniel Webster in the 19th century, and Toni Morrison and Albert Einstein in the 20th century. The current membership includes more than 250 Nobel and Pulitzer Prize winners. More

    • 88 Shares149 Views

      in Data Management & Statistics

      3 Questions: Designing software for research ethics

      by

      Data are arguably the world’s hottest form of currency, clocking in zeros and ones that hold ever more weight than before. But with all of our personal information being crunched into dynamite for enterprise solutions and the like, with a lack of consumer data protection, are we all getting left behind? 

      Jonathan Zong, a PhD candidate in electrical engineering and computer science at MIT, and an affiliate of the Computer Science and Artificial Intelligence Laboratory, thinks consent can be baked into the design of the software that gathers our data for online research. He created Bartleby, a system for debriefing research participants and eliciting their views about social media research that involved them. Using Bartleby, he says, researchers can automatically direct each of their study participants to a website where they can learn about their involvement in research, view what data researchers collected about them, and give feedback. Most importantly, participants can use the website to opt out and request to delete their data.  

      Zong and his co-author, Nathan Matias SM ’13, PhD ’17, evaluated Bartleby by debriefing thousands of participants in observational and experimental studies on Twitter and Reddit. They found that Bartleby addresses procedural concerns by creating opportunities for participants to exercise autonomy, and the tool enabled substantive, value-driven conversations about participant voice and power. Here, Zong discusses the implications of their recent work as well as the future of social, ethical, and responsible computing.

      Q: Many leading tech ethicists and policymakers believe it’s impossible to keep people informed about their involvement in research and how their data are used. How has your work changed that?

      A: When Congress asked Mark Zuckerberg in 2018 about Facebook’s obligations to keep users informed about how their data is used, his answer was effectively that all users had the opportunity to read the privacy policy, and that being any clearer would be too difficult. Tech elites often blanket-statement that ethics is complicated, and proceed with their objective anyway. Many have claimed it’s impossible to fulfill ethical responsibilities to users at scale, so why try? But by creating Bartleby, a system for debriefing participants and eliciting their views about studies that involved them, we built something that shows that it’s not only very possible, but actually pretty easy to do. In a lot of situations, letting people know we want their data and explaining why we think it’s worth it is the bare minimum we could be doing.

      Q: Can ethical challenges be solved with a software tool?

      A: Off-the-shelf software actually can make a meaningful difference in respecting people’s autonomy. Ethics regulations almost never require a debriefing process for online studies. But because we used Bartleby, people had a chance to make an informed decision. It’s a chance they otherwise wouldn’t have had.

      At the same time, we realized that using Bartleby shined a light on deeper ethics questions that required substantive reflection. For example, most people are just trying to go about their lives and ignore the messages we send them, while others reply with concerns that aren’t even always about the research. Even if indirectly, these instances help signal nuances that research participants care about.

      Where might our values as researchers differ from participants’ values? How do the power structures that shape researchers’ interaction with users and communities affect our ability to see those differences? Using software to deliver ethics procedures helps bring these questions to light. But rather than expecting definitive answers that work in every situation, we should be thinking about how using software to create opportunities for participant voice and power challenges and invites us to reflect on how we address conflicting values.

      Q: How does your approach to design help suggest a way forward for social, ethical, and responsible computing?

      A: In addition to presenting the software tool, our peer-reviewed article on Bartleby also demonstrates a theoretical framework for data ethics, inspired by ideas in feminist philosophy. Because my work spans software design, empirical social science, and philosophy, I often think about the things I want people to take away in terms of interdisciplinary bridges I want to build. 

      I hope people look at Bartleby and see that ethics is an exciting area for technical innovation that can be tested empirically — guided by a clear-headed understanding of values. Umberto Eco, a philosopher, wrote that “form must not be a vehicle for thought, it must be a way of thinking.” In other words, designing software isn’t just about putting ideas we’ve already had into a computational form. Design is also a way we can think new ideas into existence, produce new ways of knowing and doing, and imagine alternative futures. 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

    • 100 Shares149 Views

      in Data Management & Statistics

      Does this artificial intelligence think like a human?

      by

      In machine learning, understanding why a model makes certain decisions is often just as important as whether those decisions are correct. For instance, a machine-learning model might correctly predict that a skin lesion is cancerous, but it could have done so using an unrelated blip on a clinical photo.

      While tools exist to help experts make sense of a model’s reasoning, often these methods only provide insights on one decision at a time, and each must be manually evaluated. Models are commonly trained using millions of data inputs, making it almost impossible for a human to evaluate enough decisions to identify patterns.

      Now, researchers at MIT and IBM Research have created a method that enables a user to aggregate, sort, and rank these individual explanations to rapidly analyze a machine-learning model’s behavior. Their technique, called Shared Interest, incorporates quantifiable metrics that compare how well a model’s reasoning matches that of a human.

      Shared Interest could help a user easily uncover concerning trends in a model’s decision-making — for example, perhaps the model often becomes confused by distracting, irrelevant features, like background objects in photos. Aggregating these insights could help the user quickly and quantitatively determine whether a model is trustworthy and ready to be deployed in a real-world situation.

      “In developing Shared Interest, our goal is to be able to scale up this analysis process so that you could understand on a more global level what your model’s behavior is,” says lead author Angie Boggust, a graduate student in the Visualization Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      Boggust wrote the paper with her advisor, Arvind Satyanarayan, an assistant professor of computer science who leads the Visualization Group, as well as Benjamin Hoover and senior author Hendrik Strobelt, both of IBM Research. The paper will be presented at the Conference on Human Factors in Computing Systems.

      Boggust began working on this project during a summer internship at IBM, under the mentorship of Strobelt. After returning to MIT, Boggust and Satyanarayan expanded on the project and continued the collaboration with Strobelt and Hoover, who helped deploy the case studies that show how the technique could be used in practice.

      Human-AI alignment

      Shared Interest leverages popular techniques that show how a machine-learning model made a specific decision, known as saliency methods. If the model is classifying images, saliency methods highlight areas of an image that are important to the model when it made its decision. These areas are visualized as a type of heatmap, called a saliency map, that is often overlaid on the original image. If the model classified the image as a dog, and the dog’s head is highlighted, that means those pixels were important to the model when it decided the image contains a dog.

      Shared Interest works by comparing saliency methods to ground-truth data. In an image dataset, ground-truth data are typically human-generated annotations that surround the relevant parts of each image. In the previous example, the box would surround the entire dog in the photo. When evaluating an image classification model, Shared Interest compares the model-generated saliency data and the human-generated ground-truth data for the same image to see how well they align.

      The technique uses several metrics to quantify that alignment (or misalignment) and then sorts a particular decision into one of eight categories. The categories run the gamut from perfectly human-aligned (the model makes a correct prediction and the highlighted area in the saliency map is identical to the human-generated box) to completely distracted (the model makes an incorrect prediction and does not use any image features found in the human-generated box).

      “On one end of the spectrum, your model made the decision for the exact same reason a human did, and on the other end of the spectrum, your model and the human are making this decision for totally different reasons. By quantifying that for all the images in your dataset, you can use that quantification to sort through them,” Boggust explains.

      The technique works similarly with text-based data, where key words are highlighted instead of image regions.

      Rapid analysis

      The researchers used three case studies to show how Shared Interest could be useful to both nonexperts and machine-learning researchers.

      In the first case study, they used Shared Interest to help a dermatologist determine if he should trust a machine-learning model designed to help diagnose cancer from photos of skin lesions. Shared Interest enabled the dermatologist to quickly see examples of the model’s correct and incorrect predictions. Ultimately, the dermatologist decided he could not trust the model because it made too many predictions based on image artifacts, rather than actual lesions.

      “The value here is that using Shared Interest, we are able to see these patterns emerge in our model’s behavior. In about half an hour, the dermatologist was able to make a confident decision of whether or not to trust the model and whether or not to deploy it,” Boggust says.

      In the second case study, they worked with a machine-learning researcher to show how Shared Interest can evaluate a particular saliency method by revealing previously unknown pitfalls in the model. Their technique enabled the researcher to analyze thousands of correct and incorrect decisions in a fraction of the time required by typical manual methods.

      In the third case study, they used Shared Interest to dive deeper into a specific image classification example. By manipulating the ground-truth area of the image, they were able to conduct a what-if analysis to see which image features were most important for particular predictions.   

      The researchers were impressed by how well Shared Interest performed in these case studies, but Boggust cautions that the technique is only as good as the saliency methods it is based upon. If those techniques contain bias or are inaccurate, then Shared Interest will inherit those limitations.

      In the future, the researchers want to apply Shared Interest to different types of data, particularly tabular data which is used in medical records. They also want to use Shared Interest to help improve current saliency techniques. Boggust hopes this research inspires more work that seeks to quantify machine-learning model behavior in ways that make sense to humans.

      This work is funded, in part, by the MIT-IBM Watson AI Lab, the United States Air Force Research Laboratory, and the United States Air Force Artificial Intelligence Accelerator. More