Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 75 Shares99 Views

      in Data Management & Statistics

      Avoiding shortcut solutions in artificial intelligence

      by

      If your Uber driver takes a shortcut, you might get to your destination faster. But if a machine learning model takes a shortcut, it might fail in unexpected ways.

      In machine learning, a shortcut solution occurs when the model relies on a simple characteristic of a dataset to make a decision, rather than learning the true essence of the data, which can lead to inaccurate predictions. For example, a model might learn to identify images of cows by focusing on the green grass that appears in the photos, rather than the more complex shapes and patterns of the cows.  

      A new study by researchers at MIT explores the problem of shortcuts in a popular machine-learning method and proposes a solution that can prevent shortcuts by forcing the model to use more data in its decision-making.

      By removing the simpler characteristics the model is focusing on, the researchers force it to focus on more complex features of the data that it hadn’t been considering. Then, by asking the model to solve the same task two ways — once using those simpler features, and then also using the complex features it has now learned to identify — they reduce the tendency for shortcut solutions and boost the performance of the model.

      One potential application of this work is to enhance the effectiveness of machine learning models that are used to identify disease in medical images. Shortcut solutions in this context could lead to false diagnoses and have dangerous implications for patients.

      “It is still difficult to tell why deep networks make the decisions that they do, and in particular, which parts of the data these networks choose to focus upon when making a decision. If we can understand how shortcuts work in further detail, we can go even farther to answer some of the fundamental but very practical questions that are really important to people who are trying to deploy these networks,” says Joshua Robinson, a PhD student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and lead author of the paper.

      Robinson wrote the paper with his advisors, senior author Suvrit Sra, the Esther and Harold E. Edgerton Career Development Associate Professor in the Department of Electrical Engineering and Computer Science (EECS) and a core member of the Institute for Data, Systems, and Society (IDSS) and the Laboratory for Information and Decision Systems; and Stefanie Jegelka, the X-Consortium Career Development Associate Professor in EECS and a member of CSAIL and IDSS; as well as University of Pittsburgh assistant professor Kayhan Batmanghelich and PhD students Li Sun and Ke Yu. The research will be presented at the Conference on Neural Information Processing Systems in December. 

      The long road to understanding shortcuts

      The researchers focused their study on contrastive learning, which is a powerful form of self-supervised machine learning. In self-supervised machine learning, a model is trained using raw data that do not have label descriptions from humans. It can therefore be used successfully for a larger variety of data.

      A self-supervised learning model learns useful representations of data, which are used as inputs for different tasks, like image classification. But if the model takes shortcuts and fails to capture important information, these tasks won’t be able to use that information either.

      For example, if a self-supervised learning model is trained to classify pneumonia in X-rays from a number of hospitals, but it learns to make predictions based on a tag that identifies the hospital the scan came from (because some hospitals have more pneumonia cases than others), the model won’t perform well when it is given data from a new hospital.     

      For contrastive learning models, an encoder algorithm is trained to discriminate between pairs of similar inputs and pairs of dissimilar inputs. This process encodes rich and complex data, like images, in a way that the contrastive learning model can interpret.

      The researchers tested contrastive learning encoders with a series of images and found that, during this training procedure, they also fall prey to shortcut solutions. The encoders tend to focus on the simplest features of an image to decide which pairs of inputs are similar and which are dissimilar. Ideally, the encoder should focus on all the useful characteristics of the data when making a decision, Jegelka says.

      So, the team made it harder to tell the difference between the similar and dissimilar pairs, and found that this changes which features the encoder will look at to make a decision.

      “If you make the task of discriminating between similar and dissimilar items harder and harder, then your system is forced to learn more meaningful information in the data, because without learning that it cannot solve the task,” she says.

      But increasing this difficulty resulted in a tradeoff — the encoder got better at focusing on some features of the data but became worse at focusing on others. It almost seemed to forget the simpler features, Robinson says.

      To avoid this tradeoff, the researchers asked the encoder to discriminate between the pairs the same way it had originally, using the simpler features, and also after the researchers removed the information it had already learned. Solving the task both ways simultaneously caused the encoder to improve across all features.

      Their method, called implicit feature modification, adaptively modifies samples to remove the simpler features the encoder is using to discriminate between the pairs. The technique does not rely on human input, which is important because real-world data sets can have hundreds of different features that could combine in complex ways, Sra explains.

      From cars to COPD

      The researchers ran one test of this method using images of vehicles. They used implicit feature modification to adjust the color, orientation, and vehicle type to make it harder for the encoder to discriminate between similar and dissimilar pairs of images. The encoder improved its accuracy across all three features — texture, shape, and color — simultaneously.

      To see if the method would stand up to more complex data, the researchers also tested it with samples from a medical image database of chronic obstructive pulmonary disease (COPD). Again, the method led to simultaneous improvements across all features they evaluated.

      While this work takes some important steps forward in understanding the causes of shortcut solutions and working to solve them, the researchers say that continuing to refine these methods and applying them to other types of self-supervised learning will be key to future advancements.

      “This ties into some of the biggest questions about deep learning systems, like ‘Why do they fail?’ and ‘Can we know in advance the situations where your model will fail?’ There is still a lot farther to go if you want to understand shortcut learning in its full generality,” Robinson says.

      This research is supported by the National Science Foundation, National Institutes of Health, and the Pennsylvania Department of Health’s SAP SE Commonwealth Universal Research Enhancement (CURE) program. More

    • 50 Shares149 Views

      in Data Management & Statistics

      One autonomous taxi, please

      by

      If you don’t get seasick, an autonomous boat might be the right mode of transportation for you. 

      Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Senseable City Laboratory, together with Amsterdam Institute for Advanced Metropolitan Solutions (AMS Institute) in the Netherlands, have now created the final project in their self-navigating trilogy: a full-scale, fully autonomous robotic boat that’s ready to be deployed along the canals of Amsterdam. 

      “Roboat” has come a long way since the team first started prototyping small vessels in the MIT pool in late 2015. Last year, the team released their half-scale, medium model that was 2 meters long and demonstrated promising navigational prowess. 

      This year, two full-scale Roboats were launched, proving more than just proof-of-concept: these craft can comfortably carry up to five people, collect waste, deliver goods, and provide on-demand infrastructure. 

      The boat looks futuristic — it’s a sleek combination of black and gray with two seats that face each other, with orange block letters on the sides that illustrate the makers’ namesakes. It’s a fully electrical boat with a battery that’s the size of a small chest, enabling up to 10 hours of operation and wireless charging capabilities. 

      Play video

      Autonomous Roboats set sea in the Amsterdam canals and can comfortably carry up to five people, collect waste, deliver goods, and provide on-demand infrastructure.

      “We now have higher precision and robustness in the perception, navigation, and control systems, including new functions, such as close-proximity approach mode for latching capabilities, and improved dynamic positioning, so the boat can navigate real-world waters,” says Daniela Rus, MIT professor of electrical engineering and computer science and director of CSAIL. “Roboat’s control system is adaptive to the number of people in the boat.” 

      To swiftly navigate the bustling waters of Amsterdam, Roboat needs a meticulous fusion of proper navigation, perception, and control software. 

      Using GPS, the boat autonomously decides on a safe route from A to B, while continuously scanning the environment to  avoid collisions with objects, such as bridges, pillars, and other boats.

      To autonomously determine a free path and avoid crashing into objects, Roboat uses lidar and a number of cameras to enable a 360-degree view. This bundle of sensors is referred to as the “perception kit” and lets Roboat understand its surroundings. When the perception picks up an unseen object, like a canoe, for example, the algorithm flags the item as “unknown.” When the team later looks at the collected data from the day, the object is manually selected and can be tagged as “canoe.” 

      The control algorithms — similar to ones used for self-driving cars — function a little like a coxswain giving orders to rowers, by translating a given path into instructions toward the “thrusters,” which are the propellers that help the boat move.  

      If you think the boat feels slightly futuristic, its latching mechanism is one of its most impressive feats: small cameras on the boat guide it to the docking station, or other boats, when they detect specific QR codes. “The system allows Roboat to connect to other boats, and to the docking station, to form temporary bridges to alleviate traffic, as well as floating stages and squares, which wasn’t possible with the last iteration,” says Carlo Ratti, professor of the practice in the MIT Department of Urban Studies and Planning (DUSP) and director of the Senseable City Lab. 

      Roboat, by design, is also versatile. The team created a universal “hull” design — that’s the part of the boat that rides both in and on top of the water. While regular boats have unique hulls, designed for specific purposes, Roboat has a universal hull design where the base is the same, but the top decks can be switched out depending on the use case.

      “As Roboat can perform its tasks 24/7, and without a skipper on board, it adds great value for a city. However, for safety reasons it is questionable if reaching level A autonomy is desirable,” says Fabio Duarte, a principal research scientist in DUSP and lead scientist on the project. “Just like a bridge keeper, an onshore operator will monitor Roboat remotely from a control center. One operator can monitor over 50 Roboat units, ensuring smooth operations.”

      The next step for Roboat is to pilot the technology in the public domain. “The historic center of Amsterdam is the perfect place to start, with its capillary network of canals suffering from contemporary challenges, such as mobility and logistics,” says Stephan van Dijk, director of innovation at AMS Institute. 

      Previous iterations of Roboat have been presented at the IEEE International Conference on Robotics and Automation. The boats will be unveiled on Oct. 28 in the waters of Amsterdam. 

      Ratti, Rus, Duarte, and Dijk worked on the project alongside Andrew Whittle, MIT’s Edmund K Turner Professor in civil and environmental engineering; Dennis Frenchman, professor at MIT’s Department of Urban Studies and Planning; and Ynse Deinema of AMS Institute. The full team can be found at Roboat’s website. The project is a joint collaboration with AMS Institute. The City of Amsterdam is a project partner. More

    • 113 Shares189 Views

      in Data Management & Statistics

      Deep learning helps predict traffic crashes before they happen

      by

      Today’s world is one big maze, connected by layers of concrete and asphalt that afford us the luxury of navigation by vehicle. For many of our road-related advancements — GPS lets us fire fewer neurons thanks to map apps, cameras alert us to potentially costly scrapes and scratches, and electric autonomous cars have lower fuel costs — our safety measures haven’t quite caught up. We still rely on a steady diet of traffic signals, trust, and the steel surrounding us to safely get from point A to point B. 

      To get ahead of the uncertainty inherent to crashes, scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Center for Artificial Intelligence developed a deep learning model that predicts very high-resolution crash risk maps. Fed on a combination of historical crash data, road maps, satellite imagery, and GPS traces, the risk maps describe the expected number of crashes over a period of time in the future, to identify high-risk areas and predict future crashes. 

      Typically, these types of risk maps are captured at much lower resolutions that hover around hundreds of meters, which means glossing over crucial details since the roads become blurred together. These maps, though, are 5×5 meter grid cells, and the higher resolution brings newfound clarity: The scientists found that a highway road, for example, has a higher risk than nearby residential roads, and ramps merging and exiting the highway have an even higher risk than other roads. 

      “By capturing the underlying risk distribution that determines the probability of future crashes at all places, and without any historical data, we can find safer routes, enable auto insurance companies to provide customized insurance plans based on driving trajectories of customers, help city planners design safer roads, and even predict future crashes,” says MIT CSAIL PhD student Songtao He, a lead author on a new paper about the research. 

      Even though car crashes are sparse, they cost about 3 percent of the world’s GDP and are the leading cause of death in children and young adults. This sparsity makes inferring maps at such a high resolution a tricky task. Crashes at this level are thinly scattered — the average annual odds of a crash in a 5×5 grid cell is about one-in-1,000 — and they rarely happen at the same location twice. Previous attempts to predict crash risk have been largely “historical,” as an area would only be considered high-risk if there was a previous nearby crash. 

      The team’s approach casts a wider net to capture critical data. It identifies high-risk locations using GPS trajectory patterns, which give information about density, speed, and direction of traffic, and satellite imagery that describes road structures, such as the number of lanes, whether there’s a shoulder, or if there’s a large number of pedestrians. Then, even if a high-risk area has no recorded crashes, it can still be identified as high-risk, based on its traffic patterns and topology alone. 

      To evaluate the model, the scientists used crashes and data from 2017 and 2018, and tested its performance at predicting crashes in 2019 and 2020. Many locations were identified as high-risk, even though they had no recorded crashes, and also experienced crashes during the follow-up years.

      “Our model can generalize from one city to another by combining multiple clues from seemingly unrelated data sources. This is a step toward general AI, because our model can predict crash maps in uncharted territories,” says Amin Sadeghi, a lead scientist at Qatar Computing Research Institute (QCRI) and an author on the paper. “The model can be used to infer a useful crash map even in the absence of historical crash data, which could translate to positive use for city planning and policymaking by comparing imaginary scenarios.” 

      The dataset covered 7,500 square kilometers from Los Angeles, New York City, Chicago and Boston. Among the four cities, L.A. was the most unsafe, since it had the highest crash density, followed by New York City, Chicago, and Boston. 

      “If people can use the risk map to identify potentially high-risk road segments, they can take action in advance to reduce the risk of trips they take. Apps like Waze and Apple Maps have incident feature tools, but we’re trying to get ahead of the crashes — before they happen,” says He. 

      He and Sadeghi wrote the paper alongside Sanjay Chawla, research director at QCRI, and MIT professors of electrical engineering and computer science Mohammad Alizadeh, ​​Hari Balakrishnan, and Sam Madden. They will present the paper at the 2021 International Conference on Computer Vision. More

    • 63 Shares159 Views

      in Data Management & Statistics

      Making data visualizations more accessible

      by

      In the early days of the Covid-19 pandemic, the Centers for Disease Control and Prevention produced a simple chart to illustrate how measures like mask wearing and social distancing could “flatten the curve” and reduce the peak of infections.

      The chart was amplified by news sites and shared on social media platforms, but it often lacked a corresponding text description to make it accessible for blind individuals who use a screen reader to navigate the web, shutting out many of the 253 million people worldwide who have visual disabilities.

      This alternative text is often missing from online charts, and even when it is included, it is frequently uninformative or even incorrect, according to qualitative data gathered by scientists at MIT.

      These researchers conducted a study with blind and sighted readers to determine which text is useful to include in a chart description, which text is not, and why. Ultimately, they found that captions for blind readers should focus on the overall trends and statistics in the chart, not its design elements or higher-level insights.

      They also created a conceptual model that can be used to evaluate a chart description, whether the text was generated automatically by software or manually by a human author. Their work could help journalists, academics, and communicators create descriptions that are more effective for blind individuals and guide researchers as they develop better tools to automatically generate captions.

      “Ninety-nine-point-nine percent of images on Twitter lack any kind of description — and that is not hyperbole, that is the actual statistic,” says Alan Lundgard, a graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and lead author of the paper. “Having people manually author those descriptions seems to be difficult for a variety of reasons. Perhaps semiautonomous tools could help with that. But it is crucial to do this preliminary participatory design work to figure out what is the target for these tools, so we are not generating content that is either not useful to its intended audience or, in the worst case, erroneous.”

      Lundgard wrote the paper with senior author Arvind Satyanarayan, an assistant professor of computer science who leads the Visualization Group in CSAIL. The research will be presented at the Institute of Electrical and Electronics Engineers Visualization Conference in October.

      Evaluating visualizations

      To develop the conceptual model, the researchers planned to begin by studying graphs featured by popular online publications such as FiveThirtyEight and NYTimes.com, but they ran into a problem — those charts mostly lacked any textual descriptions. So instead, they collected descriptions for these charts from graduate students in an MIT data visualization class and through an online survey, then grouped the captions into four categories.

      Level 1 descriptions focus on the elements of the chart, such as its title, legend, and colors. Level 2 descriptions describe statistical content, like the minimum, maximum, or correlations. Level 3 descriptions cover perceptual interpretations of the data, like complex trends or clusters. Level 4 descriptions include subjective interpretations that go beyond the data and draw on the author’s knowledge.

      In a study with blind and sighted readers, the researchers presented visualizations with descriptions at different levels and asked participants to rate how useful they were. While both groups agreed that level 1 content on its own was not very helpful, sighted readers gave level 4 content the highest marks while blind readers ranked that content among the least useful.

      Survey results revealed that a majority of blind readers were emphatic that descriptions should not contain an author’s editorialization, but rather stick to straight facts about the data. On the other hand, most sighted readers preferred a description that told a story about the data.

      “For me, a surprising finding about the lack of utility for the highest-level content is that it ties very closely to feelings about agency and control as a disabled person. In our research, blind readers specifically didn’t want the descriptions to tell them what to think about the data. They want the data to be accessible in a way that allows them to interpret it for themselves, and they want to have the agency to do that interpretation,” Lundgard says.

      A more inclusive future

      This work could have implications as data scientists continue to develop and refine machine learning methods for autogenerating captions and alternative text.

      “We are not able to do it yet, but it is not inconceivable to imagine that in the future we would be able to automate the creation of some of this higher-level content and build models that target level 2 or level 3 in our framework. And now we know what the research questions are. If we want to produce these automated captions, what should those captions say? We are able to be a bit more directed in our future research because we have these four levels,” Satyanarayan says.

      In the future, the four-level framework could also help researchers develop machine learning models that can automatically suggest effective visualizations as part of the data analysis process, or models that can extract the most useful information from a chart.

      This research could also inform future work in Satyanarayan’s group that seeks to make interactive visualizations more accessible for blind readers who use a screen reader to access and interpret the information. 

      “The question of how to ensure that charts and graphs are accessible to screen reader users is both a socially important equity issue and a challenge that can advance the state-of-the-art in AI,” says Meredith Ringel Morris, director and principal scientist of the People + AI Research team at Google Research, who was not involved with this study. “By introducing a framework for conceptualizing natural language descriptions of information graphics that is grounded in end-user needs, this work helps ensure that future AI researchers will focus their efforts on problems aligned with end-users’ values.”

      Morris adds: “Rich natural-language descriptions of data graphics will not only expand access to critical information for people who are blind, but will also benefit a much wider audience as eyes-free interactions via smart speakers, chatbots, and other AI-powered agents become increasingly commonplace.”

      This research was supported by the National Science Foundation. More

    • 100 Shares99 Views

      in Data Management & Statistics

      Enabling AI-driven health advances without sacrificing patient privacy

      by

      There’s a lot of excitement at the intersection of artificial intelligence and health care. AI has already been used to improve disease treatment and detection, discover promising new drugs, identify links between genes and diseases, and more.

      By analyzing large datasets and finding patterns, virtually any new algorithm has the potential to help patients — AI researchers just need access to the right data to train and test those algorithms. Hospitals, understandably, are hesitant to share sensitive patient information with research teams. When they do share data, it’s difficult to verify that researchers are only using the data they need and deleting it after they’re done.

      Secure AI Labs (SAIL) is addressing those problems with a technology that lets AI algorithms run on encrypted datasets that never leave the data owner’s system. Health care organizations can control how their datasets are used, while researchers can protect the confidentiality of their models and search queries. Neither party needs to see the data or the model to collaborate.

      SAIL’s platform can also combine data from multiple sources, creating rich insights that fuel more effective algorithms.

      “You shouldn’t have to schmooze with hospital executives for five years before you can run your machine learning algorithm,” says SAIL co-founder and MIT Professor Manolis Kellis, who co-founded the company with CEO Anne Kim ’16, SM ’17. “Our goal is to help patients, to help machine learning scientists, and to create new therapeutics. We want new algorithms — the best algorithms — to be applied to the biggest possible data set.”

      SAIL has already partnered with hospitals and life science companies to unlock anonymized data for researchers. In the next year, the company hopes to be working with about half of the top 50 academic medical centers in the country.

      Unleashing AI’s full potential

      As an undergraduate at MIT studying computer science and molecular biology, Kim worked with researchers in the Computer Science and Artificial Intelligence Laboratory (CSAIL) to analyze data from clinical trials, gene association studies, hospital intensive care units, and more.

      “I realized there is something severely broken in data sharing, whether it was hospitals using hard drives, ancient file transfer protocol, or even sending stuff in the mail,” Kim says. “It was all just not well-tracked.”

      Kellis, who is also a member of the Broad Institute of MIT and Harvard, has spent years establishing partnerships with hospitals and consortia across a range of diseases including cancers, heart disease, schizophrenia, and obesity. He knew that smaller research teams would struggle to get access to the same data his lab was working with.

      In 2017, Kellis and Kim decided to commercialize technology they were developing to allow AI algorithms to run on encrypted data.

      In the summer of 2018, Kim participated in the delta v startup accelerator run by the Martin Trust Center for MIT Entrepreneurship. The founders also received support from the Sandbox Innovation Fund and the Venture Mentoring Service, and made various early connections through their MIT network.

      To participate in SAIL’s program, hospitals and other health care organizations make parts of their data available to researchers by setting up a node behind their firewall. SAIL then sends encrypted algorithms to the servers where the datasets reside in a process called federated learning. The algorithms crunch the data locally in each server and transmit the results back to a central model, which updates itself. No one — not the researchers, the data owners, or even SAIL —has access to the models or the datasets.

      The approach allows a much broader set of researchers to apply their models to large datasets. To further engage the research community, Kellis’ lab at MIT has begun holding competitions in which it gives access to datasets in areas like protein function and gene expression, and challenges researchers to predict results.

      “We invite machine learning researchers to come and train on last year’s data and predict this year’s data,” says Kellis. “If we see there’s a new type of algorithm that is performing best in these community-level assessments, people can adopt it locally at many different institutions and level the playing field. So, the only thing that matters is the quality of your algorithm rather than the power of your connections.”

      By enabling a large number of datasets to be anonymized into aggregate insights, SAIL’s technology also allows researchers to study rare diseases, in which small pools of relevant patient data are often spread out among many institutions. That has historically made the data difficult to apply AI models to.

      “We’re hoping that all of these datasets will eventually be open,” Kellis says. “We can cut across all the silos and enable a new era where every patient with every rare disorder across the entire world can come together in a single keystroke to analyze data.”

      Enabling the medicine of the future

      To work with large amounts of data around specific diseases, SAIL has increasingly sought to partner with patient associations and consortia of health care groups, including an international health care consulting company and the Kidney Cancer Association. The partnerships also align SAIL with patients, the group they’re most trying to help.

      Overall, the founders are happy to see SAIL solving problems they faced in their labs for researchers around the world.

      “The right place to solve this is not an academic project. The right place to solve this is in industry, where we can provide a platform not just for my lab but for any researcher,” Kellis says. “It’s about creating an ecosystem of academia, researchers, pharma, biotech, and hospital partners. I think it’s the blending all of these different areas that will make that vision of medicine of the future become a reality.” More

    • 63 Shares149 Views

      in Data Management & Statistics

      How quickly do algorithms improve?

      by

      Algorithms are sort of like a parent to a computer. They tell the computer how to make sense of information so they can, in turn, make something useful out of it.

      The more efficient the algorithm, the less work the computer has to do. For all of the technological progress in computing hardware, and the much debated lifespan of Moore’s Law, computer performance is only one side of the picture.

      Behind the scenes a second trend is happening: Algorithms are being improved, so in turn less computing power is needed. While algorithmic efficiency may have less of a spotlight, you’d definitely notice if your trusty search engine suddenly became one-tenth as fast, or if moving through big datasets felt like wading through sludge.

      This led scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) to ask: How quickly do algorithms improve?  

      Existing data on this question were largely anecdotal, consisting of case studies of particular algorithms that were assumed to be representative of the broader scope. Faced with this dearth of evidence, the team set off to crunch data from 57 textbooks and more than 1,110 research papers, to trace the history of when algorithms got better. Some of the research papers directly reported how good new algorithms were, and others needed to be reconstructed by the authors using “pseudocode,” shorthand versions of the algorithm that describe the basic details.

      In total, the team looked at 113 “algorithm families,” sets of algorithms solving the same problem that had been highlighted as most important by computer science textbooks. For each of the 113, the team reconstructed its history, tracking each time a new algorithm was proposed for the problem and making special note of those that were more efficient. Ranging in performance and separated by decades, starting from the 1940s to now, the team found an average of eight algorithms per family, of which a couple improved its efficiency. To share this assembled database of knowledge, the team also created Algorithm-Wiki.org.

      The scientists charted how quickly these families had improved, focusing on the most-analyzed feature of the algorithms — how fast they could guarantee to solve the problem (in computer speak: “worst-case time complexity”). What emerged was enormous variability, but also important insights on how transformative algorithmic improvement has been for computer science.

      For large computing problems, 43 percent of algorithm families had year-on-year improvements that were equal to or larger than the much-touted gains from Moore’s Law. In 14 percent of problems, the improvement to performance from algorithms vastly outpaced those that have come from improved hardware. The gains from algorithm improvement were particularly large for big-data problems, so the importance of those advancements has grown in recent decades.

      The single biggest change that the authors observed came when an algorithm family transitioned from exponential to polynomial complexity. The amount of effort it takes to solve an exponential problem is like a person trying to guess a combination on a lock. If you only have a single 10-digit dial, the task is easy. With four dials like a bicycle lock, it’s hard enough that no one steals your bike, but still conceivable that you could try every combination. With 50, it’s almost impossible — it would take too many steps. Problems that have exponential complexity are like that for computers: As they get bigger they quickly outpace the ability of the computer to handle them. Finding a polynomial algorithm often solves that, making it possible to tackle problems in a way that no amount of hardware improvement can.

      As rumblings of Moore’s Law coming to an end rapidly permeate global conversations, the researchers say that computing users will increasingly need to turn to areas like algorithms for performance improvements. The team says the findings confirm that historically, the gains from algorithms have been enormous, so the potential is there. But if gains come from algorithms instead of hardware, they’ll look different. Hardware improvement from Moore’s Law happens smoothly over time, and for algorithms the gains come in steps that are usually large but infrequent. 

      “This is the first paper to show how fast algorithms are improving across a broad range of examples,” says Neil Thompson, an MIT research scientist at CSAIL and the Sloan School of Management and senior author on the new paper. “Through our analysis, we were able to say how many more tasks could be done using the same amount of computing power after an algorithm improved. As problems increase to billions or trillions of data points, algorithmic improvement becomes substantially more important than hardware improvement. In an era where the environmental footprint of computing is increasingly worrisome, this is a way to improve businesses and other organizations without the downside.”

      Thompson wrote the paper alongside MIT visiting student Yash Sherry. The paper is published in the Proceedings of the IEEE. The work was funded by the Tides foundation and the MIT Initiative on the Digital Economy. More

    • 138 Shares189 Views

      in Data Management & Statistics

      Exact symbolic artificial intelligence for faster, better assessment of AI fairness

      by

      The justice system, banks, and private companies use algorithms to make decisions that have profound impacts on people’s lives. Unfortunately, those algorithms are sometimes biased — disproportionately impacting people of color as well as individuals in lower income classes when they apply for loans or jobs, or even when courts decide what bail should be set while a person awaits trial.

      MIT researchers have developed a new artificial intelligence programming language that can assess the fairness of algorithms more exactly, and more quickly, than available alternatives.

      Their Sum-Product Probabilistic Language (SPPL) is a probabilistic programming system. Probabilistic programming is an emerging field at the intersection of programming languages and artificial intelligence that aims to make AI systems much easier to develop, with early successes in computer vision, common-sense data cleaning, and automated data modeling. Probabilistic programming languages make it much easier for programmers to define probabilistic models and carry out probabilistic inference — that is, work backward to infer probable explanations for observed data.

      “There are previous systems that can solve various fairness questions. Our system is not the first; but because our system is specialized and optimized for a certain class of models, it can deliver solutions thousands of times faster,” says Feras Saad, a PhD student in electrical engineering and computer science (EECS) and first author on a recent paper describing the work. Saad adds that the speedups are not insignificant: The system can be up to 3,000 times faster than previous approaches.

      SPPL gives fast, exact solutions to probabilistic inference questions such as “How likely is the model to recommend a loan to someone over age 40?” or “Generate 1,000 synthetic loan applicants, all under age 30, whose loans will be approved.” These inference results are based on SPPL programs that encode probabilistic models of what kinds of applicants are likely, a priori, and also how to classify them. Fairness questions that SPPL can answer include “Is there a difference between the probability of recommending a loan to an immigrant and nonimmigrant applicant with the same socioeconomic status?” or “What’s the probability of a hire, given that the candidate is qualified for the job and from an underrepresented group?”

      SPPL is different from most probabilistic programming languages, as SPPL only allows users to write probabilistic programs for which it can automatically deliver exact probabilistic inference results. SPPL also makes it possible for users to check how fast inference will be, and therefore avoid writing slow programs. In contrast, other probabilistic programming languages such as Gen and Pyro allow users to write down probabilistic programs where the only known ways to do inference are approximate — that is, the results include errors whose nature and magnitude can be hard to characterize.

      Error from approximate probabilistic inference is tolerable in many AI applications. But it is undesirable to have inference errors corrupting results in socially impactful applications of AI, such as automated decision-making, and especially in fairness analysis.

      Jean-Baptiste Tristan, associate professor at Boston College and former research scientist at Oracle Labs, who was not involved in the new research, says, “I’ve worked on fairness analysis in academia and in real-world, large-scale industry settings. SPPL offers improved flexibility and trustworthiness over other PPLs on this challenging and important class of problems due to the expressiveness of the language, its precise and simple semantics, and the speed and soundness of the exact symbolic inference engine.”

      SPPL avoids errors by restricting to a carefully designed class of models that still includes a broad class of AI algorithms, including the decision tree classifiers that are widely used for algorithmic decision-making. SPPL works by compiling probabilistic programs into a specialized data structure called a “sum-product expression.” SPPL further builds on the emerging theme of using probabilistic circuits as a representation that enables efficient probabilistic inference. This approach extends prior work on sum-product networks to models and queries expressed via a probabilistic programming language. However, Saad notes that this approach comes with limitations: “SPPL is substantially faster for analyzing the fairness of a decision tree, for example, but it can’t analyze models like neural networks. Other systems can analyze both neural networks and decision trees, but they tend to be slower and give inexact answers.”

      “SPPL shows that exact probabilistic inference is practical, not just theoretically possible, for a broad class of probabilistic programs,” says Vikash Mansinghka, an MIT principal research scientist and senior author on the paper. “In my lab, we’ve seen symbolic inference driving speed and accuracy improvements in other inference tasks that we previously approached via approximate Monte Carlo and deep learning algorithms. We’ve also been applying SPPL to probabilistic programs learned from real-world databases, to quantify the probability of rare events, generate synthetic proxy data given constraints, and automatically screen data for probable anomalies.”

      The new SPPL probabilistic programming language was presented in June at the ACM SIGPLAN International Conference on Programming Language Design and Implementation (PLDI), in a paper that Saad co-authored with MIT EECS Professor Martin Rinard and Mansinghka. SPPL is implemented in Python and is available open source. More

    • 38 Shares129 Views

      in Artificial Intelligence

      Contact-aware robot design

      by

      Adequate biomimicry in robotics necessitates a delicate balance between design and control, an integral part of making our machines more like us. Advanced dexterity in humans is wrapped up in a long evolutionary tale of how our fists of fury evolved to accomplish complex tasks. With machines, designing a new robotic manipulator could mean long, manual iteration cycles of designing, fabricating, and evaluating guided by human intuition. 

      Most robotic hands are designed for general purposes, as it’s very tedious to make task-specific hands. Existing methods battle trade-offs between the complexity of designs critical for contact-rich tasks, and the practical constraints of manufacturing, and contact handling. 

      This led researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) to create a new method to computationally optimize the shape and control of a robotic manipulator for a specific task. Their system uses software to manipulate the design, simulate the robot doing a task, and then provide an optimization score to assess the design and control. 

      Such task-driven manipulator optimization has potential for a wide range of applications in manufacturing and warehouse robot systems, where each task needs to be performed repeatedly, but different manipulators would be suitable for individual tasks. 

      Play video

      A new method to represent robotic manipulators helps optimize complex and organic shapes for future machines.

      Seeking to test the functionality of the system, the team first created a single robotic finger design to flip over a box on the ground. The fingertip structure, which looked something like Captain Hook’s left hand, was automatically optimized by an algorithm to hook onto the box’s back surface and flip it. They also developed a model for an assembly task, where a two-finger design put a small cube into a larger, movable mount. Since the fingers were two different lengths, they could reach two objects of different sizes, and the larger and flatter surfaces of the fingers helped stably push the object. 

      Traditionally, this joint optimization process consists of using simple, more primitive shapes to approximate each component of a robot design. When creating a three-segment robotic finger, for example, it would likely be approximated by three connected cylinders, where the algorithm optimizes the length and radius to achieve the desired design and shape. While this would simplify the optimization problem, oversimplifying the shape would be limiting for more complex designs, and ultimately complex tasks. 

      To create more involved manipulators, the team’s method used a technique called “cage-based deformation,” which essentially lets the user change or deform the geometry of a shape in real-time.

      Using the software, you’d put something that looks like a cage around the robotic finger, for example. The algorithm can automatically change the cage dimensions to make more sophisticated, natural shapes. The different variations of designs still keep their integrity, so they can be easily fabricated.

      A simulator was developed by the team to simulate the manipulator design and control on a task, which then provides a performance score.

      “Using these simulation tools, we don’t need to evaluate the design by manufacturing and testing it in the real world,” says Jie Xu, MIT PhD student and lead author on a new paper about the research. “In contrast to reinforcement learning algorithms that are popular for manipulation, but are data-inefficient, the proposed cage-based representation and the simulator allows for the use of powerful gradient-based methods. We not only find better solutions, but also find them faster. As a result we can quickly score the design, thus significantly shortening the design cycle.”

      In the future, the team plans to extend the software to optimize the manipulators concurrently for multiple tasks.

      Xu wrote the paper alongside MIT PhD student Tao Chen, MIT graduate student Lara Zlokapa, MIT research scientist Michael Foshey, MIT Professor Wojciech Matusik, Texas A&M University Assistant professor Shinjiro Sueda, and MIT Professor Pulkit Agrawal. They presented the paper virtually at the 2021 Robotic Science and Systems conference last week. The work is supported by the Toyota Research Institute. More