Machine learning

Subterms

    More stories

    • 113 Shares99 Views

      in Data Management & Statistics

      Researchers use large language models to help robots navigate

      by

      Someday, you may want your home robot to carry a load of dirty clothes downstairs and deposit them in the washing machine in the far-left corner of the basement. The robot will need to combine your instructions with its visual observations to determine the steps it should take to complete this task.For an AI agent, this is easier said than done. Current approaches often utilize multiple hand-crafted machine-learning models to tackle different parts of the task, which require a great deal of human effort and expertise to build. These methods, which use visual representations to directly make navigation decisions, demand massive amounts of visual data for training, which are often hard to come by.To overcome these challenges, researchers from MIT and the MIT-IBM Watson AI Lab devised a navigation method that converts visual representations into pieces of language, which are then fed into one large language model that achieves all parts of the multistep navigation task.Rather than encoding visual features from images of a robot’s surroundings as visual representations, which is computationally intensive, their method creates text captions that describe the robot’s point-of-view. A large language model uses the captions to predict the actions a robot should take to fulfill a user’s language-based instructions.Because their method utilizes purely language-based representations, they can use a large language model to efficiently generate a huge amount of synthetic training data.While this approach does not outperform techniques that use visual features, it performs well in situations that lack enough visual data for training. The researchers found that combining their language-based inputs with visual signals leads to better navigation performance.“By purely using language as the perceptual representation, ours is a more straightforward approach. Since all the inputs can be encoded as language, we can generate a human-understandable trajectory,” says Bowen Pan, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on this approach.Pan’s co-authors include his advisor, Aude Oliva, director of strategic industry engagement at the MIT Schwarzman College of Computing, MIT director of the MIT-IBM Watson AI Lab, and a senior research scientist in the Computer Science and Artificial Intelligence Laboratory (CSAIL); Philip Isola, an associate professor of EECS and a member of CSAIL; senior author Yoon Kim, an assistant professor of EECS and a member of CSAIL; and others at the MIT-IBM Watson AI Lab and Dartmouth College. The research will be presented at the Conference of the North American Chapter of the Association for Computational Linguistics.Solving a vision problem with languageSince large language models are the most powerful machine-learning models available, the researchers sought to incorporate them into the complex task known as vision-and-language navigation, Pan says.But such models take text-based inputs and can’t process visual data from a robot’s camera. So, the team needed to find a way to use language instead.Their technique utilizes a simple captioning model to obtain text descriptions of a robot’s visual observations. These captions are combined with language-based instructions and fed into a large language model, which decides what navigation step the robot should take next.The large language model outputs a caption of the scene the robot should see after completing that step. This is used to update the trajectory history so the robot can keep track of where it has been.The model repeats these processes to generate a trajectory that guides the robot to its goal, one step at a time.To streamline the process, the researchers designed templates so observation information is presented to the model in a standard form — as a series of choices the robot can make based on its surroundings.For instance, a caption might say “to your 30-degree left is a door with a potted plant beside it, to your back is a small office with a desk and a computer,” etc. The model chooses whether the robot should move toward the door or the office.“One of the biggest challenges was figuring out how to encode this kind of information into language in a proper way to make the agent understand what the task is and how they should respond,” Pan says.Advantages of languageWhen they tested this approach, while it could not outperform vision-based techniques, they found that it offered several advantages.First, because text requires fewer computational resources to synthesize than complex image data, their method can be used to rapidly generate synthetic training data. In one test, they generated 10,000 synthetic trajectories based on 10 real-world, visual trajectories.The technique can also bridge the gap that can prevent an agent trained with a simulated environment from performing well in the real world. This gap often occurs because computer-generated images can appear quite different from real-world scenes due to elements like lighting or color. But language that describes a synthetic versus a real image would be much harder to tell apart, Pan says. Also, the representations their model uses are easier for a human to understand because they are written in natural language.“If the agent fails to reach its goal, we can more easily determine where it failed and why it failed. Maybe the history information is not clear enough or the observation ignores some important details,” Pan says.In addition, their method could be applied more easily to varied tasks and environments because it uses only one type of input. As long as data can be encoded as language, they can use the same model without making any modifications.But one disadvantage is that their method naturally loses some information that would be captured by vision-based models, such as depth information.However, the researchers were surprised to see that combining language-based representations with vision-based methods improves an agent’s ability to navigate.“Maybe this means that language can capture some higher-level information than cannot be captured with pure vision features,” he says.This is one area the researchers want to continue exploring. They also want to develop a navigation-oriented captioner that could boost the method’s performance. In addition, they want to probe the ability of large language models to exhibit spatial awareness and see how this could aid language-based navigation.This research is funded, in part, by the MIT-IBM Watson AI Lab. More

    • 88 Shares119 Views

      in Data Management & Statistics

      Making climate models relevant for local decision-makers

      by

      Climate models are a key technology in predicting the impacts of climate change. By running simulations of the Earth’s climate, scientists and policymakers can estimate conditions like sea level rise, flooding, and rising temperatures, and make decisions about how to appropriately respond. But current climate models struggle to provide this information quickly or affordably enough to be useful on smaller scales, such as the size of a city. Now, authors of a new open-access paper published in the Journal of Advances in Modeling Earth Systems have found a method to leverage machine learning to utilize the benefits of current climate models, while reducing the computational costs needed to run them. “It turns the traditional wisdom on its head,” says Sai Ravela, a principal research scientist in MIT’s Department of Earth, Atmospheric and Planetary Sciences (EAPS) who wrote the paper with EAPS postdoc Anamitra Saha. Traditional wisdomIn climate modeling, downscaling is the process of using a global climate model with coarse resolution to generate finer details over smaller regions. Imagine a digital picture: A global model is a large picture of the world with a low number of pixels. To downscale, you zoom in on just the section of the photo you want to look at — for example, Boston. But because the original picture was low resolution, the new version is blurry; it doesn’t give enough detail to be particularly useful. “If you go from coarse resolution to fine resolution, you have to add information somehow,” explains Saha. Downscaling attempts to add that information back in by filling in the missing pixels. “That addition of information can happen two ways: Either it can come from theory, or it can come from data.” Conventional downscaling often involves using models built on physics (such as the process of air rising, cooling, and condensing, or the landscape of the area), and supplementing it with statistical data taken from historical observations. But this method is computationally taxing: It takes a lot of time and computing power to run, while also being expensive. A little bit of both In their new paper, Saha and Ravela have figured out a way to add the data another way. They’ve employed a technique in machine learning called adversarial learning. It uses two machines: One generates data to go into our photo. But the other machine judges the sample by comparing it to actual data. If it thinks the image is fake, then the first machine has to try again until it convinces the second machine. The end-goal of the process is to create super-resolution data. Using machine learning techniques like adversarial learning is not a new idea in climate modeling; where it currently struggles is its inability to handle large amounts of basic physics, like conservation laws. The researchers discovered that simplifying the physics going in and supplementing it with statistics from the historical data was enough to generate the results they needed. “If you augment machine learning with some information from the statistics and simplified physics both, then suddenly, it’s magical,” says Ravela. He and Saha started with estimating extreme rainfall amounts by removing more complex physics equations and focusing on water vapor and land topography. They then generated general rainfall patterns for mountainous Denver and flat Chicago alike, applying historical accounts to correct the output. “It’s giving us extremes, like the physics does, at a much lower cost. And it’s giving us similar speeds to statistics, but at much higher resolution.” Another unexpected benefit of the results was how little training data was needed. “The fact that that only a little bit of physics and little bit of statistics was enough to improve the performance of the ML [machine learning] model … was actually not obvious from the beginning,” says Saha. It only takes a few hours to train, and can produce results in minutes, an improvement over the months other models take to run. Quantifying risk quicklyBeing able to run the models quickly and often is a key requirement for stakeholders such as insurance companies and local policymakers. Ravela gives the example of Bangladesh: By seeing how extreme weather events will impact the country, decisions about what crops should be grown or where populations should migrate to can be made considering a very broad range of conditions and uncertainties as soon as possible.“We can’t wait months or years to be able to quantify this risk,” he says. “You need to look out way into the future and at a large number of uncertainties to be able to say what might be a good decision.”While the current model only looks at extreme precipitation, training it to examine other critical events, such as tropical storms, winds, and temperature, is the next step of the project. With a more robust model, Ravela is hoping to apply it to other places like Boston and Puerto Rico as part of a Climate Grand Challenges project.“We’re very excited both by the methodology that we put together, as well as the potential applications that it could lead to,” he says.  More

    • 88 Shares149 Views

      in Data Management & Statistics

      A data-driven approach to making better choices

      by

      Imagine a world in which some important decision — a judge’s sentencing recommendation, a child’s treatment protocol, which person or business should receive a loan — was made more reliable because a well-designed algorithm helped a key decision-maker arrive at a better choice. A new MIT economics course is investigating these interesting possibilities.Class 14.163 (Algorithms and Behavioral Science) is a new cross-disciplinary course focused on behavioral economics, which studies the cognitive capacities and limitations of human beings. The course was co-taught this past spring by assistant professor of economics Ashesh Rambachan and visiting lecturer Sendhil Mullainathan.Rambachan studies the economic applications of machine learning, focusing on algorithmic tools that drive decision-making in the criminal justice system and consumer lending markets. He also develops methods for determining causation using cross-sectional and dynamic data.Mullainathan will soon join the MIT departments of Electrical Engineering and Computer Science and Economics as a professor. His research uses machine learning to understand complex problems in human behavior, social policy, and medicine. Mullainathan co-founded the Abdul Latif Jameel Poverty Action Lab (J-PAL) in 2003.The new course’s goals are both scientific (to understand people) and policy-driven (to improve society by improving decisions). Rambachan believes that machine-learning algorithms provide new tools for both the scientific and applied goals of behavioral economics.“The course investigates the deployment of computer science, artificial intelligence (AI), economics, and machine learning in service of improved outcomes and reduced instances of bias in decision-making,” Rambachan says.There are opportunities, Rambachan believes, for constantly evolving digital tools like AI, machine learning, and large language models (LLMs) to help reshape everything from discriminatory practices in criminal sentencing to health-care outcomes among underserved populations.Students learn how to use machine learning tools with three main objectives: to understand what they do and how they do it, to formalize behavioral economics insights so they compose well within machine learning tools, and to understand areas and topics where the integration of behavioral economics and algorithmic tools might be most fruitful.Students also produce ideas, develop associated research, and see the bigger picture. They’re led to understand where an insight fits and see where the broader research agenda is leading. Participants can think critically about what supervised LLMs can (and cannot) do, to understand how to integrate those capacities with the models and insights of behavioral economics, and to recognize the most fruitful areas for the application of what investigations uncover.The dangers of subjectivity and biasAccording to Rambachan, behavioral economics acknowledges that biases and mistakes exist throughout our choices, even absent algorithms. “The data used by our algorithms exist outside computer science and machine learning, and instead are often produced by people,” he continues. “Understanding behavioral economics is therefore essential to understanding the effects of algorithms and how to better build them.”Rambachan sought to make the course accessible regardless of attendees’ academic backgrounds. The class included advanced degree students from a variety of disciplines.By offering students a cross-disciplinary, data-driven approach to investigating and discovering ways in which algorithms might improve problem-solving and decision-making, Rambachan hopes to build a foundation on which to redesign existing systems of jurisprudence, health care, consumer lending, and industry, to name a few areas.“Understanding how data are generated can help us understand bias,” Rambachan says. “We can ask questions about producing a better outcome than what currently exists.”Useful tools for re-imagining social operationsEconomics doctoral student Jimmy Lin was skeptical about the claims Rambachan and Mullainathan made when the class began, but changed his mind as the course continued.“Ashesh and Sendhil started with two provocative claims: The future of behavioral science research will not exist without AI, and the future of AI research will not exist without behavioral science,” Lin says. “Over the course of the semester, they deepened my understanding of both fields and walked us through numerous examples of how economics informed AI research and vice versa.”Lin, who’d previously done research in computational biology, praised the instructors’ emphasis on the importance of a “producer mindset,” thinking about the next decade of research rather than the previous decade. “That’s especially important in an area as interdisciplinary and fast-moving as the intersection of AI and economics — there isn’t an old established literature, so you’re forced to ask new questions, invent new methods, and create new bridges,” he says.The speed of change to which Lin alludes is a draw for him, too. “We’re seeing black-box AI methods facilitate breakthroughs in math, biology, physics, and other scientific disciplines,” Lin  says. “AI can change the way we approach intellectual discovery as researchers.”An interdisciplinary future for economics and social systemsStudying traditional economic tools and enhancing their value with AI may yield game-changing shifts in how institutions and organizations teach and empower leaders to make choices.“We’re learning to track shifts, to adjust frameworks and better understand how to deploy tools in service of a common language,” Rambachan says. “We must continually interrogate the intersection of human judgment, algorithms, AI, machine learning, and LLMs.”Lin enthusiastically recommended the course regardless of students’ backgrounds. “Anyone broadly interested in algorithms in society, applications of AI across academic disciplines, or AI as a paradigm for scientific discovery should take this class,” he says. “Every lecture felt like a goldmine of perspectives on research, novel application areas, and inspiration on how to produce new, exciting ideas.”The course, Rambachan says, argues that better-built algorithms can improve decision-making across disciplines. “By building connections between economics, computer science, and machine learning, perhaps we can automate the best of human choices to improve outcomes while minimizing or eliminating the worst,” he says.Lin remains excited about the course’s as-yet unexplored possibilities. “It’s a class that makes you excited about the future of research and your own role in it,” he says. More

    • 88 Shares129 Views

      in Data Management & Statistics

      A technique for more effective multipurpose robots

      by

      Let’s say you want to train a robot so it understands how to use tools and can then quickly learn to make repairs around your house with a hammer, wrench, and screwdriver. To do that, you would need an enormous amount of data demonstrating tool use.Existing robotic datasets vary widely in modality — some include color images while others are composed of tactile imprints, for instance. Data could also be collected in different domains, like simulation or human demos. And each dataset may capture a unique task and environment.It is difficult to efficiently incorporate data from so many sources in one machine-learning model, so many methods use just one type of data to train a robot. But robots trained this way, with a relatively small amount of task-specific data, are often unable to perform new tasks in unfamiliar environments.In an effort to train better multipurpose robots, MIT researchers developed a technique to combine multiple sources of data across domains, modalities, and tasks using a type of generative AI known as diffusion models.They train a separate diffusion model to learn a strategy, or policy, for completing one task using one specific dataset. Then they combine the policies learned by the diffusion models into a general policy that enables a robot to perform multiple tasks in various settings.In simulations and real-world experiments, this training approach enabled a robot to perform multiple tool-use tasks and adapt to new tasks it did not see during training. The method, known as Policy Composition (PoCo), led to a 20 percent improvement in task performance when compared to baseline techniques.“Addressing heterogeneity in robotic datasets is like a chicken-egg problem. If we want to use a lot of data to train general robot policies, then we first need deployable robots to get all this data. I think that leveraging all the heterogeneous data available, similar to what researchers have done with ChatGPT, is an important step for the robotics field,” says Lirui Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on PoCo.     Wang’s coauthors include Jialiang Zhao, a mechanical engineering graduate student; Yilun Du, an EECS graduate student; Edward Adelson, the John and Dorothy Wilson Professor of Vision Science in the Department of Brain and Cognitive Sciences and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering, and a member of CSAIL. The research will be presented at the Robotics: Science and Systems Conference.Combining disparate datasetsA robotic policy is a machine-learning model that takes inputs and uses them to perform an action. One way to think about a policy is as a strategy. In the case of a robotic arm, that strategy might be a trajectory, or a series of poses that move the arm so it picks up a hammer and uses it to pound a nail.Datasets used to learn robotic policies are typically small and focused on one particular task and environment, like packing items into boxes in a warehouse.“Every single robotic warehouse is generating terabytes of data, but it only belongs to that specific robot installation working on those packages. It is not ideal if you want to use all of these data to train a general machine,” Wang says.The MIT researchers developed a technique that can take a series of smaller datasets, like those gathered from many robotic warehouses, learn separate policies from each one, and combine the policies in a way that enables a robot to generalize to many tasks.They represent each policy using a type of generative AI model known as a diffusion model. Diffusion models, often used for image generation, learn to create new data samples that resemble samples in a training dataset by iteratively refining their output.But rather than teaching a diffusion model to generate images, the researchers teach it to generate a trajectory for a robot. They do this by adding noise to the trajectories in a training dataset. The diffusion model gradually removes the noise and refines its output into a trajectory.This technique, known as Diffusion Policy, was previously introduced by researchers at MIT, Columbia University, and the Toyota Research Institute. PoCo builds off this Diffusion Policy work. The team trains each diffusion model with a different type of dataset, such as one with human video demonstrations and another gleaned from teleoperation of a robotic arm.Then the researchers perform a weighted combination of the individual policies learned by all the diffusion models, iteratively refining the output so the combined policy satisfies the objectives of each individual policy.Greater than the sum of its parts“One of the benefits of this approach is that we can combine policies to get the best of both worlds. For instance, a policy trained on real-world data might be able to achieve more dexterity, while a policy trained on simulation might be able to achieve more generalization,” Wang says.

      With policy composition, researchers are able to combine datasets from multiple sources so they can teach a robot to effectively use a wide range of tools, like a hammer, screwdriver, or this spatula.Image: Courtesy of the researchers

      Because the policies are trained separately, one could mix and match diffusion policies to achieve better results for a certain task. A user could also add data in a new modality or domain by training an additional Diffusion Policy with that dataset, rather than starting the entire process from scratch.

      The policy composition technique the researchers developed can be used to effectively teach a robot to use tools even when objects are placed around it to try and distract it from its task, as seen here.Image: Courtesy of the researchers

      The researchers tested PoCo in simulation and on real robotic arms that performed a variety of tools tasks, such as using a hammer to pound a nail and flipping an object with a spatula. PoCo led to a 20 percent improvement in task performance compared to baseline methods.“The striking thing was that when we finished tuning and visualized it, we can clearly see that the composed trajectory looks much better than either one of them individually,” Wang says.In the future, the researchers want to apply this technique to long-horizon tasks where a robot would pick up one tool, use it, then switch to another tool. They also want to incorporate larger robotics datasets to improve performance.“We will need all three kinds of data to succeed for robotics: internet data, simulation data, and real robot data. How to combine them effectively will be the million-dollar question. PoCo is a solid step on the right track,” says Jim Fan, senior research scientist at NVIDIA and leader of the AI Agents Initiative, who was not involved with this work.This research is funded, in part, by Amazon, the Singapore Defense Science and Technology Agency, the U.S. National Science Foundation, and the Toyota Research Institute. More

    • 200 Shares159 Views

      in Data Management & Statistics

      Looking for a specific action in a video? This AI-based method can find it for you

      by

      The internet is awash in instructional videos that can teach curious viewers everything from cooking the perfect pancake to performing a life-saving Heimlich maneuver.But pinpointing when and where a particular action happens in a long video can be tedious. To streamline the process, scientists are trying to teach computers to perform this task. Ideally, a user could just describe the action they’re looking for, and an AI model would skip to its location in the video.However, teaching machine-learning models to do this usually requires a great deal of expensive video data that have been painstakingly hand-labeled.A new, more efficient approach from researchers at MIT and the MIT-IBM Watson AI Lab trains a model to perform this task, known as spatio-temporal grounding, using only videos and their automatically generated transcripts.The researchers teach a model to understand an unlabeled video in two distinct ways: by looking at small details to figure out where objects are located (spatial information) and looking at the bigger picture to understand when the action occurs (temporal information).Compared to other AI approaches, their method more accurately identifies actions in longer videos with multiple activities. Interestingly, they found that simultaneously training on spatial and temporal information makes a model better at identifying each individually.In addition to streamlining online learning and virtual training processes, this technique could also be useful in health care settings by rapidly finding key moments in videos of diagnostic procedures, for example.“We disentangle the challenge of trying to encode spatial and temporal information all at once and instead think about it like two experts working on their own, which turns out to be a more explicit way to encode the information. Our model, which combines these two separate branches, leads to the best performance,” says Brian Chen, lead author of a paper on this technique.Chen, a 2023 graduate of Columbia University who conducted this research while a visiting student at the MIT-IBM Watson AI Lab, is joined on the paper by James Glass, senior research scientist, member of the MIT-IBM Watson AI Lab, and head of the Spoken Language Systems Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL); Hilde Kuehne, a member of the MIT-IBM Watson AI Lab who is also affiliated with Goethe University Frankfurt; and others at MIT, Goethe University, the MIT-IBM Watson AI Lab, and Quality Match GmbH. The research will be presented at the Conference on Computer Vision and Pattern Recognition.Global and local learningResearchers usually teach models to perform spatio-temporal grounding using videos in which humans have annotated the start and end times of particular tasks.Not only is generating these data expensive, but it can be difficult for humans to figure out exactly what to label. If the action is “cooking a pancake,” does that action start when the chef begins mixing the batter or when she pours it into the pan?“This time, the task may be about cooking, but next time, it might be about fixing a car. There are so many different domains for people to annotate. But if we can learn everything without labels, it is a more general solution,” Chen says.For their approach, the researchers use unlabeled instructional videos and accompanying text transcripts from a website like YouTube as training data. These don’t need any special preparation.They split the training process into two pieces. For one, they teach a machine-learning model to look at the entire video to understand what actions happen at certain times. This high-level information is called a global representation.For the second, they teach the model to focus on a specific region in parts of the video where action is happening. In a large kitchen, for instance, the model might only need to focus on the wooden spoon a chef is using to mix pancake batter, rather than the entire counter. This fine-grained information is called a local representation.The researchers incorporate an additional component into their framework to mitigate misalignments that occur between narration and video. Perhaps the chef talks about cooking the pancake first and performs the action later.To develop a more realistic solution, the researchers focused on uncut videos that are several minutes long. In contrast, most AI techniques train using few-second clips that someone trimmed to show only one action.A new benchmarkBut when they came to evaluate their approach, the researchers couldn’t find an effective benchmark for testing a model on these longer, uncut videos — so they created one.To build their benchmark dataset, the researchers devised a new annotation technique that works well for identifying multistep actions. They had users mark the intersection of objects, like the point where a knife edge cuts a tomato, rather than drawing a box around important objects.“This is more clearly defined and speeds up the annotation process, which reduces the human labor and cost,” Chen says.Plus, having multiple people do point annotation on the same video can better capture actions that occur over time, like the flow of milk being poured. All annotators won’t mark the exact same point in the flow of liquid.When they used this benchmark to test their approach, the researchers found that it was more accurate at pinpointing actions than other AI techniques.Their method was also better at focusing on human-object interactions. For instance, if the action is “serving a pancake,” many other approaches might focus only on key objects, like a stack of pancakes sitting on a counter. Instead, their method focuses on the actual moment when the chef flips a pancake onto a plate.Next, the researchers plan to enhance their approach so models can automatically detect when text and narration are not aligned, and switch focus from one modality to the other. They also want to extend their framework to audio data, since there are usually strong correlations between actions and the sounds objects make.“AI research has made incredible progress towards creating models like ChatGPT that understand images. But our progress on understanding video is far behind. This work represents a significant step forward in that direction,” says Kate Saenko, a professor in the Department of Computer Science at Boston University who was not involved with this work.This research is funded, in part, by the MIT-IBM Watson AI Lab. More

    • 88 Shares159 Views

      in Data Management & Statistics

      Exploring the mysterious alphabet of sperm whales

      by

      The allure of whales has stoked human consciousness for millennia, casting these ocean giants as enigmatic residents of the deep seas. From the biblical Leviathan to Herman Melville’s formidable Moby Dick, whales have been central to mythologies and folklore. And while cetology, or whale science, has improved our knowledge of these marine mammals in the past century in particular, studying whales has remained a formidable a challenge.Now, thanks to machine learning, we’re a little closer to understanding these gentle giants. Researchers from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and Project CETI (Cetacean Translation Initiative) recently used algorithms to decode the “sperm whale phonetic alphabet,” revealing sophisticated structures in sperm whale communication akin to human phonetics and communication systems in other animal species. In a new open-access study published in Nature Communications, the research shows that sperm whales codas, or short bursts of clicks that they use to communicate, vary significantly in structure depending on the conversational context, revealing a communication system far more intricate than previously understood. 

      Play video

      The Secret Language of Sperm Whales, DecodedVideo: MIT CSAIL

      Nine thousand codas, collected from Eastern Caribbean sperm whale families observed by the Dominica Sperm Whale Project, proved an instrumental starting point in uncovering the creatures’ complex communication system. Alongside the data gold mine, the team used a mix of algorithms for pattern recognition and classification, as well as on-body recording equipment. It turned out that sperm whale communications were indeed not random or simplistic, but rather structured in a complex, combinatorial manner. The researchers identified something of a “sperm whale phonetic alphabet,” where various elements that researchers call  “rhythm,” “tempo,” “rubato,” and “ornamentation” interplay to form a vast array of distinguishable codas. For example, the whales would systematically modulate certain aspects of their codas based on the conversational context, such as smoothly varying the duration of the calls — rubato — or adding extra ornamental clicks. But even more remarkably, they found that the basic building blocks of these codas could be combined in a combinatorial fashion, allowing the whales to construct a vast repertoire of distinct vocalizations.The experiments were conducted using acoustic bio-logging tags (specifically something called “D-tags”) deployed on whales from the Eastern Caribbean clan. These tags captured the intricate details of the whales’ vocal patterns. By developing new visualization and data analysis techniques, the CSAIL researchers found that individual sperm whales could emit various coda patterns in long exchanges, not just repeats of the same coda. These patterns, they say, are nuanced, and include fine-grained variations that other whales also produce and recognize.“We are venturing into the unknown, to decipher the mysteries of sperm whale communication without any pre-existing ground truth data,” says Daniela Rus, CSAIL director and professor of electrical engineering and computer science (EECS) at MIT. “Using machine learning is important for identifying the features of their communications and predicting what they say next. Our findings indicate the presence of structured information content and also challenges the prevailing belief among many linguists that complex communication is unique to humans. This is a step toward showing that other species have levels of communication complexity that have not been identified so far, deeply connected to behavior. Our next steps aim to decipher the meaning behind these communications and explore the societal-level correlations between what is being said and group actions.”Whaling aroundSperm whales have the largest brains among all known animals. This is accompanied by very complex social behaviors between families and cultural groups, necessitating strong communication for coordination, especially in pressurized environments like deep sea hunting.Whales owe much to Roger Payne, former Project CETI advisor, whale biologist, conservationist, and MacArthur Fellow who was a major figure in elucidating their musical careers. In the noted 1971 Science article “Songs of Humpback Whales,” Payne documented how whales can sing. His work later catalyzed the “Save the Whales” movement, a successful and timely conservation initiative.“Roger’s research highlights the impact science can have on society. His finding that whales sing led to the marine mammal protection act and helped save several whale species from extinction. This interdisciplinary research now brings us one step closer to knowing what sperm whales are saying,” says David Gruber, lead and founder of Project CETI and distinguished professor of biology at the City University of New York.Today, CETI’s upcoming research aims to discern whether elements like rhythm, tempo, ornamentation, and rubato carry specific communicative intents, potentially providing insights into the “duality of patterning” — a linguistic phenomenon where simple elements combine to convey complex meanings previously thought unique to human language.Aliens among us“One of the intriguing aspects of our research is that it parallels the hypothetical scenario of contacting alien species. It’s about understanding a species with a completely different environment and communication protocols, where their interactions are distinctly different from human norms,” says Pratyusha Sharma, an MIT PhD student in EECS, CSAIL affiliate, and the study’s lead author. “We’re exploring how to interpret the basic units of meaning in their communication. This isn’t just about teaching animals a subset of human language, but decoding a naturally evolved communication system within their unique biological and environmental constraints. Essentially, our work could lay the groundwork for deciphering how an ‘alien civilization’ might communicate, providing insights into creating algorithms or systems to understand entirely unfamiliar forms of communication.”“Many animal species have repertoires of several distinct signals, but we are only beginning to uncover the extent to which they combine these signals to create new messages,” says Robert Seyfarth, a University of Pennsylvania professor emeritus of psychology who was not involved in the research. “Scientists are particularly interested in whether signal combinations vary according to the social or ecological context in which they are given, and the extent to which signal combinations follow discernible ‘rules’ that are recognized by listeners. The problem is particularly challenging in the case of marine mammals, because scientists usually cannot see their subjects or identify in complete detail the context of communication. Nonetheless, this paper offers new, tantalizing details of call combinations and the rules that underlie them in sperm whales.”Joining Sharma, Rus, and Gruber are two others from MIT, both CSAIL principal investigators and professors in EECS: Jacob Andreas and Antonio Torralba. They join Shane Gero, biology lead at CETI, founder of the Dominica Sperm Whale Project, and scientist-in residence at Carleton University. The paper was funded by Project CETI via Dalio Philanthropies and Ocean X, Sea Grape Foundation, Rosamund Zander/Hansjorg Wyss, and Chris Anderson/Jacqueline Novogratz through The Audacious Project: a collaborative funding initiative housed at TED, with further support from the J.H. and E.V. Wade Fund at MIT. More

    • 113 Shares149 Views

      in Data Management & Statistics

      An AI dataset carves new paths to tornado detection

      by

      The return of spring in the Northern Hemisphere touches off tornado season. A tornado’s twisting funnel of dust and debris seems an unmistakable sight. But that sight can be obscured to radar, the tool of meteorologists. It’s hard to know exactly when a tornado has formed, or even why.

      A new dataset could hold answers. It contains radar returns from thousands of tornadoes that have hit the United States in the past 10 years. Storms that spawned tornadoes are flanked by other severe storms, some with nearly identical conditions, that never did. MIT Lincoln Laboratory researchers who curated the dataset, called TorNet, have now released it open source. They hope to enable breakthroughs in detecting one of nature’s most mysterious and violent phenomena.

      “A lot of progress is driven by easily available, benchmark datasets. We hope TorNet will lay a foundation for machine learning algorithms to both detect and predict tornadoes,” says Mark Veillette, the project’s co-principal investigator with James Kurdzo. Both researchers work in the Air Traffic Control Systems Group. 

      Along with the dataset, the team is releasing models trained on it. The models show promise for machine learning’s ability to spot a twister. Building on this work could open new frontiers for forecasters, helping them provide more accurate warnings that might save lives. 

      Swirling uncertainty

      About 1,200 tornadoes occur in the United States every year, causing millions to billions of dollars in economic damage and claiming 71 lives on average. Last year, one unusually long-lasting tornado killed 17 people and injured at least 165 others along a 59-mile path in Mississippi.  

      Yet tornadoes are notoriously difficult to forecast because scientists don’t have a clear picture of why they form. “We can see two storms that look identical, and one will produce a tornado and one won’t. We don’t fully understand it,” Kurdzo says.

      A tornado’s basic ingredients are thunderstorms with instability caused by rapidly rising warm air and wind shear that causes rotation. Weather radar is the primary tool used to monitor these conditions. But tornadoes lay too low to be detected, even when moderately close to the radar. As the radar beam with a given tilt angle travels further from the antenna, it gets higher above the ground, mostly seeing reflections from rain and hail carried in the “mesocyclone,” the storm’s broad, rotating updraft. A mesocyclone doesn’t always produce a tornado.

      With this limited view, forecasters must decide whether or not to issue a tornado warning. They often err on the side of caution. As a result, the rate of false alarms for tornado warnings is more than 70 percent. “That can lead to boy-who-cried-wolf syndrome,” Kurdzo says.  

      In recent years, researchers have turned to machine learning to better detect and predict tornadoes. However, raw datasets and models have not always been accessible to the broader community, stifling progress. TorNet is filling this gap.

      The dataset contains more than 200,000 radar images, 13,587 of which depict tornadoes. The rest of the images are non-tornadic, taken from storms in one of two categories: randomly selected severe storms or false-alarm storms (those that led a forecaster to issue a warning but that didn’t produce a tornado).

      Each sample of a storm or tornado comprises two sets of six radar images. The two sets correspond to different radar sweep angles. The six images portray different radar data products, such as reflectivity (showing precipitation intensity) or radial velocity (indicating if winds are moving toward or away from the radar).

      A challenge in curating the dataset was first finding tornadoes. Within the corpus of weather radar data, tornadoes are extremely rare events. The team then had to balance those tornado samples with difficult non-tornado samples. If the dataset were too easy, say by comparing tornadoes to snowstorms, an algorithm trained on the data would likely over-classify storms as tornadic.

      “What’s beautiful about a true benchmark dataset is that we’re all working with the same data, with the same level of difficulty, and can compare results,” Veillette says. “It also makes meteorology more accessible to data scientists, and vice versa. It becomes easier for these two parties to work on a common problem.”

      Both researchers represent the progress that can come from cross-collaboration. Veillette is a mathematician and algorithm developer who has long been fascinated by tornadoes. Kurdzo is a meteorologist by training and a signal processing expert. In grad school, he chased tornadoes with custom-built mobile radars, collecting data to analyze in new ways.

      “This dataset also means that a grad student doesn’t have to spend a year or two building a dataset. They can jump right into their research,” Kurdzo says.

      This project was funded by Lincoln Laboratory’s Climate Change Initiative, which aims to leverage the laboratory’s diverse technical strengths to help address climate problems threatening human health and global security.

      Chasing answers with deep learning

      Using the dataset, the researchers developed baseline artificial intelligence (AI) models. They were particularly eager to apply deep learning, a form of machine learning that excels at processing visual data. On its own, deep learning can extract features (key observations that an algorithm uses to make a decision) from images across a dataset. Other machine learning approaches require humans to first manually label features. 

      “We wanted to see if deep learning could rediscover what people normally look for in tornadoes and even identify new things that typically aren’t searched for by forecasters,” Veillette says.

      The results are promising. Their deep learning model performed similar to or better than all tornado-detecting algorithms known in literature. The trained algorithm correctly classified 50 percent of weaker EF-1 tornadoes and over 85 percent of tornadoes rated EF-2 or higher, which make up the most devastating and costly occurrences of these storms.

      They also evaluated two other types of machine-learning models, and one traditional model to compare against. The source code and parameters of all these models are freely available. The models and dataset are also described in a paper submitted to a journal of the American Meteorological Society (AMS). Veillette presented this work at the AMS Annual Meeting in January.

      “The biggest reason for putting our models out there is for the community to improve upon them and do other great things,” Kurdzo says. “The best solution could be a deep learning model, or someone might find that a non-deep learning model is actually better.”

      TorNet could be useful in the weather community for others uses too, such as for conducting large-scale case studies on storms. It could also be augmented with other data sources, like satellite imagery or lightning maps. Fusing multiple types of data could improve the accuracy of machine learning models.

      Taking steps toward operations

      On top of detecting tornadoes, Kurdzo hopes that models might help unravel the science of why they form.

      “As scientists, we see all these precursors to tornadoes — an increase in low-level rotation, a hook echo in reflectivity data, specific differential phase (KDP) foot and differential reflectivity (ZDR) arcs. But how do they all go together? And are there physical manifestations we don’t know about?” he asks.

      Teasing out those answers might be possible with explainable AI. Explainable AI refers to methods that allow a model to provide its reasoning, in a format understandable to humans, of why it came to a certain decision. In this case, these explanations might reveal physical processes that happen before tornadoes. This knowledge could help train forecasters, and models, to recognize the signs sooner. 

      “None of this technology is ever meant to replace a forecaster. But perhaps someday it could guide forecasters’ eyes in complex situations, and give a visual warning to an area predicted to have tornadic activity,” Kurdzo says.

      Such assistance could be especially useful as radar technology improves and future networks potentially grow denser. Data refresh rates in a next-generation radar network are expected to increase from every five minutes to approximately one minute, perhaps faster than forecasters can interpret the new information. Because deep learning can process huge amounts of data quickly, it could be well-suited for monitoring radar returns in real time, alongside humans. Tornadoes can form and disappear in minutes.

      But the path to an operational algorithm is a long road, especially in safety-critical situations, Veillette says. “I think the forecaster community is still, understandably, skeptical of machine learning. One way to establish trust and transparency is to have public benchmark datasets like this one. It’s a first step.”

      The next steps, the team hopes, will be taken by researchers across the world who are inspired by the dataset and energized to build their own algorithms. Those algorithms will in turn go into test beds, where they’ll eventually be shown to forecasters, to start a process of transitioning into operations.

      In the end, the path could circle back to trust.

      “We may never get more than a 10- to 15-minute tornado warning using these tools. But if we could lower the false-alarm rate, we could start to make headway with public perception,” Kurdzo says. “People are going to use those warnings to take the action they need to save their lives.” More

    • 175 Shares179 Views

      in Data Management & Statistics

      This tiny chip can safeguard user data while enabling efficient computing on a smartphone

      by

      Health-monitoring apps can help people manage chronic diseases or stay on track with fitness goals, using nothing more than a smartphone. However, these apps can be slow and energy-inefficient because the vast machine-learning models that power them must be shuttled between a smartphone and a central memory server.

      Engineers often speed things up using hardware that reduces the need to move so much data back and forth. While these machine-learning accelerators can streamline computation, they are susceptible to attackers who can steal secret information.

      To reduce this vulnerability, researchers from MIT and the MIT-IBM Watson AI Lab created a machine-learning accelerator that is resistant to the two most common types of attacks. Their chip can keep a user’s health records, financial information, or other sensitive data private while still enabling huge AI models to run efficiently on devices.

      The team developed several optimizations that enable strong security while only slightly slowing the device. Moreover, the added security does not impact the accuracy of computations. This machine-learning accelerator could be particularly beneficial for demanding AI applications like augmented and virtual reality or autonomous driving.

      While implementing the chip would make a device slightly more expensive and less energy-efficient, that is sometimes a worthwhile price to pay for security, says lead author Maitreyi Ashok, an electrical engineering and computer science (EECS) graduate student at MIT.

      “It is important to design with security in mind from the ground up. If you are trying to add even a minimal amount of security after a system has been designed, it is prohibitively expensive. We were able to effectively balance a lot of these tradeoffs during the design phase,” says Ashok.

      Her co-authors include Saurav Maji, an EECS graduate student; Xin Zhang and John Cohn of the MIT-IBM Watson AI Lab; and senior author Anantha Chandrakasan, MIT’s chief innovation and strategy officer, dean of the School of Engineering, and the Vannevar Bush Professor of EECS. The research will be presented at the IEEE Custom Integrated Circuits Conference.

      Side-channel susceptibility

      The researchers targeted a type of machine-learning accelerator called digital in-memory compute. A digital IMC chip performs computations inside a device’s memory, where pieces of a machine-learning model are stored after being moved over from a central server.

      The entire model is too big to store on the device, but by breaking it into pieces and reusing those pieces as much as possible, IMC chips reduce the amount of data that must be moved back and forth.

      But IMC chips can be susceptible to hackers. In a side-channel attack, a hacker monitors the chip’s power consumption and uses statistical techniques to reverse-engineer data as the chip computes. In a bus-probing attack, the hacker can steal bits of the model and dataset by probing the communication between the accelerator and the off-chip memory.

      Digital IMC speeds computation by performing millions of operations at once, but this complexity makes it tough to prevent attacks using traditional security measures, Ashok says.

      She and her collaborators took a three-pronged approach to blocking side-channel and bus-probing attacks.

      First, they employed a security measure where data in the IMC are split into random pieces. For instance, a bit zero might be split into three bits that still equal zero after a logical operation. The IMC never computes with all pieces in the same operation, so a side-channel attack could never reconstruct the real information.

      But for this technique to work, random bits must be added to split the data. Because digital IMC performs millions of operations at once, generating so many random bits would involve too much computing. For their chip, the researchers found a way to simplify computations, making it easier to effectively split data while eliminating the need for random bits.

      Second, they prevented bus-probing attacks using a lightweight cipher that encrypts the model stored in off-chip memory. This lightweight cipher only requires simple computations. In addition, they only decrypted the pieces of the model stored on the chip when necessary.

      Third, to improve security, they generated the key that decrypts the cipher directly on the chip, rather than moving it back and forth with the model. They generated this unique key from random variations in the chip that are introduced during manufacturing, using what is known as a physically unclonable function.

      “Maybe one wire is going to be a little bit thicker than another. We can use these variations to get zeros and ones out of a circuit. For every chip, we can get a random key that should be consistent because these random properties shouldn’t change significantly over time,” Ashok explains.

      They reused the memory cells on the chip, leveraging the imperfections in these cells to generate the key. This requires less computation than generating a key from scratch.

      “As security has become a critical issue in the design of edge devices, there is a need to develop a complete system stack focusing on secure operation. This work focuses on security for machine-learning workloads and describes a digital processor that uses cross-cutting optimization. It incorporates encrypted data access between memory and processor, approaches to preventing side-channel attacks using randomization, and exploiting variability to generate unique codes. Such designs are going to be critical in future mobile devices,” says Chandrakasan.

      Safety testing

      To test their chip, the researchers took on the role of hackers and tried to steal secret information using side-channel and bus-probing attacks.

      Even after making millions of attempts, they couldn’t reconstruct any real information or extract pieces of the model or dataset. The cipher also remained unbreakable. By contrast, it took only about 5,000 samples to steal information from an unprotected chip.

      The addition of security did reduce the energy efficiency of the accelerator, and it also required a larger chip area, which would make it more expensive to fabricate.

      The team is planning to explore methods that could reduce the energy consumption and size of their chip in the future, which would make it easier to implement at scale.

      “As it becomes too expensive, it becomes harder to convince someone that security is critical. Future work could explore these tradeoffs. Maybe we could make it a little less secure but easier to implement and less expensive,” Ashok says.

      The research is funded, in part, by the MIT-IBM Watson AI Lab, the National Science Foundation, and a Mathworks Engineering Fellowship. More