Algorithms

Subterms

    More stories

    • 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

    • 138 Shares179 Views

      in Data Management & Statistics

      3 Questions: Kalyan Veeramachaneni on hurdles preventing fully automated machine learning

      by

      The proliferation of big data across domains, from banking to health care to environmental monitoring, has spurred increasing demand for machine learning tools that help organizations make decisions based on the data they gather.

      That growing industry demand has driven researchers to explore the possibilities of automated machine learning (AutoML), which seeks to automate the development of machine learning solutions in order to make them accessible for nonexperts, improve their efficiency, and accelerate machine learning research. For example, an AutoML system might enable doctors to use their expertise interpreting electroencephalography (EEG) results to build a model that can predict which patients are at higher risk for epilepsy — without requiring the doctors to have a background in data science.

      Yet, despite more than a decade of work, researchers have been unable to fully automate all steps in the machine learning development process. Even the most efficient commercial AutoML systems still require a prolonged back-and-forth between a domain expert, like a marketing manager or mechanical engineer, and a data scientist, making the process inefficient.

      Kalyan Veeramachaneni, a principal research scientist in the MIT Laboratory for Information and Decision Systems who has been studying AutoML since 2010, has co-authored a paper in the journal ACM Computing Surveys that details a seven-tiered schematic to evaluate AutoML tools based on their level of autonomy.

      A system at level zero has no automation and requires a data scientist to start from scratch and build models by hand, while a tool at level six is completely automated and can be easily and effectively used by a nonexpert. Most commercial systems fall somewhere in the middle.

      Veeramachaneni spoke with MIT News about the current state of AutoML, the hurdles that prevent truly automatic machine learning systems, and the road ahead for AutoML researchers.

      Q: How has automatic machine learning evolved over the past decade, and what is the current state of AutoML systems?

      A: In 2010, we started to see a shift, with enterprises wanting to invest in getting value out of their data beyond just business intelligence. So then came the question, maybe there are certain things in the development of machine learning-based solutions that we can automate? The first iteration of AutoML was to make our own jobs as data scientists more efficient. Can we take away the grunt work that we do on a day-to-day basis and automate that by using a software system? That area of research ran its course until about 2015, when we realized we still weren’t able to speed up this development process.

      Then another thread emerged. There are a lot of problems that could be solved with data, and they come from experts who know those problems, who live with them on a daily basis. These individuals have very little to do with machine learning or software engineering. How do we bring them into the fold? That is really the next frontier.

      There are three areas where these domain experts have strong input in a machine learning system. The first is defining the problem itself and then helping to formulate it as a prediction task to be solved by a machine learning model. Second, they know how the data have been collected, so they also know intuitively how to process that data. And then third, at the end, machine learning models only give you a very tiny part of a solution — they just give you a prediction. The output of a machine learning model is just one input to help a domain expert get to a decision or action.

      Q: What steps of the machine learning pipeline are the most difficult to automate, and why has automating them been so challenging?

      A: The problem-formulation part is extremely difficult to automate. For example, if I am a researcher who wants to get more government funding, and I have a lot of data about the content of the research proposals that I write and whether or not I receive funding, can machine learning help there? We don’t know yet. In problem formulation, I use my domain expertise to translate the problem into something that is more tangible to predict, and that requires somebody who knows the domain very well. And he or she also knows how to use that information post-prediction. That problem is refusing to be automated.

      There is one part of problem-formulation that could be automated. It turns out that we can look at the data and mathematically express several possible prediction tasks automatically. Then we can share those prediction tasks with the domain expert to see if any of them would help in the larger problem they are trying to tackle. Then once you pick the prediction task, there are a lot of intermediate steps you do, including feature engineering, modeling, etc., that are very mechanical steps and easy to automate.

      But defining the prediction tasks has typically been a collaborative effort between data scientists and domain experts because, unless you know the domain, you can’t translate the domain problem into a prediction task. And then sometimes domain experts don’t know what is meant by “prediction.” That leads to the major, significant back and forth in the process. If you automate that step, then machine learning penetration and the use of data to create meaningful predictions will increase tremendously.

      Then what happens after the machine learning model gives a prediction? We can automate the software and technology part of it, but at the end of the day, it is root cause analysis and human intuition and decision making. We can augment them with a lot of tools, but we can’t fully automate that.

      Q: What do you hope to achieve with the seven-tiered framework for evaluating AutoML systems that you outlined in your paper?

      A: My hope is that people start to recognize that some levels of automation have already been achieved and some still need to be tackled. In the research community, we tend to focus on what we are comfortable with. We have gotten used to automating certain steps, and then we just stick to it. Automating these other parts of the machine learning solution development is very important, and that is where the biggest bottlenecks remain.

      My second hope is that researchers will very clearly understand what domain expertise means. A lot of this AutoML work is still being conducted by academics, and the problem is that we often don’t do applied work. There is not a crystal-clear definition of what a domain expert is and in itself, “domain expert,” is a very nebulous phrase. What we mean by domain expert is the expert in the problem you are trying to solve with machine learning. And I am hoping that everyone unifies around that because that would make things so much clearer.

      I still believe that we are not able to build that many models for that many problems, but even for the ones that we are building, the majority of them are not getting deployed and used in day-to-day life. The output of machine learning is just going to be another data point, an augmented data point, in someone’s decision making. How they make those decisions, based on that input, how that will change their behavior, and how they will adapt their style of working, that is still a big, open question. Once we automate everything, that is what’s next.

      We have to determine what has to fundamentally change in the day-to-day workflow of someone giving loans at a bank, or an educator trying to decide whether he or she should change the assignments in an online class. How are they going to use machine learning’s outputs? We need to focus on the fundamental things we have to build out to make machine learning more usable. 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

    • 88 Shares179 Views

      in Data Management & Statistics

      Research collaboration puts climate-resilient crops in sight

      by

      Any houseplant owner knows that changes in the amount of water or sunlight a plant receives can put it under immense stress. A dying plant brings certain disappointment to anyone with a green thumb. 

      But for farmers who make their living by successfully growing plants, and whose crops may nourish hundreds or thousands of people, the devastation of failing flora is that much greater. As climate change is poised to cause increasingly unpredictable weather patterns globally, crops may be subject to more extreme environmental conditions like droughts, fluctuating temperatures, floods, and wildfire. 

      Climate scientists and food systems researchers worry about the stress climate change may put on crops, and on global food security. In an ambitious interdisciplinary project funded by the Abdul Latif Jameel Water and Food Systems Lab (J-WAFS), David Des Marais, the Gale Assistant Professor in the Department of Civil and Environmental Engineering at MIT, and Caroline Uhler, an associate professor in the MIT Department of Electrical Engineering and Computer Science and the Institute for Data, Systems, and Society, are investigating how plant genes communicate with one another under stress. Their research results can be used to breed plants more resilient to climate change.

      Crops in trouble

      Governing plants’ responses to environmental stress are gene regulatory networks, or GRNs, which guide the development and behaviors of living things. A GRN may be comprised of thousands of genes and proteins that all communicate with one another. GRNs help a particular cell, tissue, or organism respond to environmental changes by signaling certain genes to turn their expression on or off.

      Even seemingly minor or short-term changes in weather patterns can have large effects on crop yield and food security. An environmental trigger, like a lack of water during a crucial phase of plant development, can turn a gene on or off, and is likely to affect many others in the GRN. For example, without water, a gene enabling photosynthesis may switch off. This can create a domino effect, where the genes that rely on those regulating photosynthesis are silenced, and the cycle continues. As a result, when photosynthesis is halted, the plant may experience other detrimental side effects, like no longer being able to reproduce or defend against pathogens. The chain reaction could even kill a plant before it has the chance to be revived by a big rain.

      Des Marais says he wishes there was a way to stop those genes from completely shutting off in such a situation. To do that, scientists would need to better understand how exactly gene networks respond to different environmental triggers. Bringing light to this molecular process is exactly what he aims to do in this collaborative research effort.

      Solving complex problems across disciplines

      Despite their crucial importance, GRNs are difficult to study because of how complex and interconnected they are. Usually, to understand how a particular gene is affecting others, biologists must silence one gene and see how the others in the network respond. 

      For years, scientists have aspired to an algorithm that could synthesize the massive amount of information contained in GRNs to “identify correct regulatory relationships among genes,” according to a 2019 article in the Encyclopedia of Bioinformatics and Computational Biology. 

      “A GRN can be seen as a large causal network, and understanding the effects that silencing one gene has on all other genes requires understanding the causal relationships among the genes,” says Uhler. “These are exactly the kinds of algorithms my group develops.”

      Des Marais and Uhler’s project aims to unravel these complex communication networks and discover how to breed crops that are more resilient to the increased droughts, flooding, and erratic weather patterns that climate change is already causing globally.

      In addition to climate change, by 2050, the world will demand 70 percent more food to feed a booming population. “Food systems challenges cannot be addressed individually in disciplinary or topic area silos,” says Greg Sixt, J-WAFS’ research manager for climate and food systems. “They must be addressed in a systems context that reflects the interconnected nature of the food system.”

      Des Marais’ background is in biology, and Uhler’s in statistics. “Dave’s project with Caroline was essentially experimental,” says Renee J. Robins, J-WAFS’ executive director. “This kind of exploratory research is exactly what the J-WAFS seed grant program is for.”

      Getting inside gene regulatory networks

      Des Marais and Uhler’s work begins in a windowless basement on MIT’s campus, where 300 genetically identical Brachypodium distachyon plants grow in large, temperature-controlled chambers. The plant, which contains more than 30,000 genes, is a good model for studying important cereal crops like wheat, barley, maize, and millet. For three weeks, all plants receive the same temperature, humidity, light, and water. Then, half are slowly tapered off water, simulating drought-like conditions.

      Six days into the forced drought, the plants are clearly suffering. Des Marais’ PhD student Jie Yun takes tissues from 50 hydrated and 50 dry plants, freezes them in liquid nitrogen to immediately halt metabolic activity, grinds them up into a fine powder, and chemically separates the genetic material. The genes from all 100 samples are then sequenced at a lab across the street.

      The team is left with a spreadsheet listing the 30,000 genes found in each of the 100 plants at the moment they were frozen, and how many copies there were. Uhler’s PhD student Anastasiya Belyaeva inputs the massive spreadsheet into the computer program she developed and runs her novel algorithm. Within a few hours, the group can see which genes were most active in one condition over another, how the genes were communicating, and which were causing changes in others. 

      The methodology captures important subtleties that could allow researchers to eventually alter gene pathways and breed more resilient crops. “When you expose a plant to drought stress, it’s not like there’s some canonical response,” Des Marais says. “There’s lots of things going on. It’s turning this physiologic process up, this one down, this one didn’t exist before, and now suddenly is turned on.” 

      In addition to Des Marais and Uhler’s research, J-WAFS has funded projects in food and water from researchers in 29 departments across all five MIT schools as well as the MIT Schwarzman College of Computing. J-WAFS seed grants typically fund seven to eight new projects every year.

      “The grants are really aimed at catalyzing new ideas, providing the sort of support [for MIT researchers] to be pushing boundaries, and also bringing in faculty who may have some interesting ideas that they haven’t yet applied to water or food concerns,” Robins says. “It’s an avenue for researchers all over the Institute to apply their ideas to water and food.”

      Alison Gold is a student in MIT’s Graduate Program in Science Writing. More

    • 163 Shares99 Views

      in Data Management & Statistics

      A universal system for decoding any type of data sent across a network

      by

      Every piece of data that travels over the internet — from paragraphs in an email to 3D graphics in a virtual reality environment — can be altered by the noise it encounters along the way, such as electromagnetic interference from a microwave or Bluetooth device. The data are coded so that when they arrive at their destination, a decoding algorithm can undo the negative effects of that noise and retrieve the original data.

      Since the 1950s, most error-correcting codes and decoding algorithms have been designed together. Each code had a structure that corresponded with a particular, highly complex decoding algorithm, which often required the use of dedicated hardware.

      Researchers at MIT, Boston University, and Maynooth University in Ireland have now created the first silicon chip that is able to decode any code, regardless of its structure, with maximum accuracy, using a universal decoding algorithm called Guessing Random Additive Noise Decoding (GRAND). By eliminating the need for multiple, computationally complex decoders, GRAND enables increased efficiency that could have applications in augmented and virtual reality, gaming, 5G networks, and connected devices that rely on processing a high volume of data with minimal delay.

      The research at MIT is led by Muriel Médard, the Cecil H. and Ida Green Professor in the Department of Electrical Engineering and Computer Science, and was co-authored by Amit Solomon and Wei Ann, both graduate students at MIT; Rabia Tugce Yazicigil, assistant professor of electrical and computer engineering at Boston University; Arslan Riaz and Vaibhav Bansal, both graduate students at Boston University; Ken R. Duffy, director of the Hamilton Institute at the National University of Ireland at Maynooth; and Kevin Galligan, a Maynooth graduate student. The research will be presented at the European Solid-States Device Research and Circuits Conference next week.

      Focus on noise

      One way to think of these codes is as redundant hashes (in this case, a series of 1s and 0s) added to the end of the original data. The rules for the creation of that hash are stored in a specific codebook.

      As the encoded data travel over a network, they are affected by noise, or energy that disrupts the signal, which is often generated by other electronic devices. When that coded data and the noise that affected them arrive at their destination, the decoding algorithm consults its codebook and uses the structure of the hash to guess what the stored information is.

      Instead, GRAND works by guessing the noise that affected the message, and uses the noise pattern to deduce the original information. GRAND generates a series of noise sequences in the order they are likely to occur, subtracts them from the received data, and checks to see if the resulting codeword is in a codebook.

      While the noise appears random in nature, it has a probabilistic structure that allows the algorithm to guess what it might be.

      “In a way, it is similar to troubleshooting. If someone brings their car into the shop, the mechanic doesn’t start by mapping the entire car to blueprints. Instead, they start by asking, ‘What is the most likely thing to go wrong?’ Maybe it just needs gas. If that doesn’t work, what’s next? Maybe the battery is dead?” Médard says.

      Novel hardware

      The GRAND chip uses a three-tiered structure, starting with the simplest possible solutions in the first stage and working up to longer and more complex noise patterns in the two subsequent stages. Each stage operates independently, which increases the throughput of the system and saves power.

      The device is also designed to switch seamlessly between two codebooks. It contains two static random-access memory chips, one that can crack codewords, while the other loads a new codebook and then switches to decoding without any downtime.

      The researchers tested the GRAND chip and found it could effectively decode any moderate redundancy code up to 128 bits in length, with only about a microsecond of latency.

      Médard and her collaborators had previously demonstrated the success of the algorithm, but this new work showcases the effectiveness and efficiency of GRAND in hardware for the first time.

      Developing hardware for the novel decoding algorithm required the researchers to first toss aside their preconceived notions, Médard says.

      “We couldn’t go out and reuse things that had already been done. This was like a complete whiteboard. We had to really think about every single component from scratch. It was a journey of reconsideration. And I think when we do our next chip, there will be things with this first chip that we’ll realize we did out of habit or assumption that we can do better,” she says.

      A chip for the future

      Since GRAND only uses codebooks for verification, the chip not only works with legacy codes but could also be used with codes that haven’t even been introduced yet.

      In the lead-up to 5G implementation, regulators and communications companies struggled to find consensus as to which codes should be used in the new network. Regulators ultimately chose to use two types of traditional codes for 5G infrastructure in different situations. Using GRAND could eliminate the need for that rigid standardization in the future, Médard says.

      The GRAND chip could even open the field of coding to a wave of innovation.

      “For reasons I’m not quite sure of, people approach coding with awe, like it is black magic. The process is mathematically nasty, so people just use codes that already exist. I’m hoping this will recast the discussion so it is not so standards-oriented, enabling people to use codes that already exist and create new codes,” she says.

      Moving forward, Médard and her collaborators plan to tackle the problem of soft detection with a retooled version of the GRAND chip. In soft detection, the received data are less precise.

      They also plan to test the ability of GRAND to crack longer, more complex codes and adjust the structure of the silicon chip to improve its energy efficiency.

      The research was funded by the Battelle Memorial Institute and Science Foundation of Ireland. More

    • 100 Shares149 Views

      in Data Management & Statistics

      Using adversarial attacks to refine molecular energy predictions

      by

      Neural networks (NNs) are increasingly being used to predict new materials, the rate and yield of chemical reactions, and drug-target interactions, among others. For these applications, they are orders of magnitude faster than traditional methods such as quantum mechanical simulations. 

      The price for this agility, however, is reliability. Because machine learning models only interpolate, they may fail when used outside the domain of training data.

      But the part that worried Rafael Gómez-Bombarelli, the Jeffrey Cheah Career Development Professor in the MIT Department of Materials Science and Engineering, and graduate students Daniel Schwalbe-Koda and Aik Rui Tan was that establishing the limits of these machine learning (ML) models is tedious and labor-intensive. 

      This is particularly true for predicting ‘‘potential energy surfaces” (PES), or the map of a molecule’s energy in all its configurations. These surfaces encode the complexities of a molecule into flatlands, valleys, peaks, troughs, and ravines. The most stable configurations of a system are usually in the deep pits — quantum mechanical chasms from which atoms and molecules typically do not escape. 

      In a recent Nature Communications paper, the research team presented a way to demarcate the “safe zone” of a neural network by using “adversarial attacks.” Adversarial attacks have been studied for other classes of problems, such as image classification, but this is the first time that they are being used to sample molecular geometries in a PES. 

      “People have been using uncertainty for active learning for years in ML potentials. The key difference is that they need to run the full ML simulation and evaluate if the NN was reliable, and if it wasn’t, acquire more data, retrain and re-simulate. Meaning that it takes a long time to nail down the right model, and one has to run the ML simulation many times” explains Gómez-Bombarelli.

      The Gómez-Bombarelli lab at MIT works on a synergistic synthesis of first-principles simulation and machine learning that greatly speeds up this process. The actual simulations are run only for a small fraction of these molecules, and all those data are fed into a neural network that learns how to predict the same properties for the rest of the molecules. They have successfully demonstrated these methods for a growing class of novel materials that includes catalysts for producing hydrogen from water, cheaper polymer electrolytes for electric vehicles,  zeolites for molecular sieving, magnetic materials, and more. 

      The challenge, however, is that these neural networks are only as smart as the data they are trained on.  Considering the PES map, 99 percent of the data may fall into one pit, totally missing valleys that are of more interest. 

      Such wrong predictions can have disastrous consequences — think of a self-driving car that fails to identify a person crossing the street.

      One way to find out the uncertainty of a model is to run the same data through multiple versions of it. 

      For this project, the researchers had multiple neural networks predict the potential energy surface from the same data. Where the network is fairly sure of the prediction, the variation between the outputs of different networks is minimal and the surfaces largely converge. When the network is uncertain, the predictions of different models vary widely, producing a range of outputs, any of which could be the correct surface. 

      The spread in the predictions of a “committee of neural networks” is the “uncertainty” at that point. A good model should not just indicate the best prediction, but also indicates the uncertainty about each of these predictions. It’s like the neural network says “this property for material A will have a value of X and I’m highly confident about it.”

      This could have been an elegant solution but for the sheer scale of the combinatorial space. “Each simulation (which is ground feed for the neural network) may take from tens to thousands of CPU hours,” explains Schwalbe-Koda. For the results to be meaningful, multiple models must be run over a sufficient number of points in the PES, an extremely time-consuming process. 

      Instead, the new approach only samples data points from regions of low prediction confidence, corresponding to specific geometries of a molecule. These molecules are then stretched or deformed slightly so that the uncertainty of the neural network committee is maximized. Additional data are computed for these molecules through simulations and then added to the initial training pool. 

      The neural networks are trained again, and a new set of uncertainties are calculated. This process is repeated until the uncertainty associated with various points on the surface becomes well-defined and cannot be decreased any further. 

      Gómez-Bombarelli explains, “We aspire to have a model that is perfect in the regions we care about (i.e., the ones that the simulation will visit) without having had to run the full ML simulation, by making sure that we make it very good in high-likelihood regions where it isn’t.”

      The paper presents several examples of this approach, including predicting complex supramolecular interactions in zeolites. These materials are cavernous crystals that act as molecular sieves with high shape selectivity. They find applications in catalysis, gas separation, and ion exchange, among others.

      Because performing simulations of large zeolite structures is very costly, the researchers show how their method can provide significant savings in computational simulations. They used more than 15,000 examples to train a neural network to predict the potential energy surfaces for these systems. Despite the large cost required to generate the dataset, the final results are mediocre, with only around 80 percent of the neural network-based simulations being successful. To improve the performance of the model using traditional active learning methods, the researchers calculated an additional 5,000 data points, which improved the performance of the neural network potentials to 92 percent.

      However, when the adversarial approach is used to retrain the neural networks, the authors saw a performance jump to 97 percent using only 500 extra points. That’s a remarkable result, the researchers say, especially considering that each of these extra points takes hundreds of CPU hours. 

      This could be the most realistic method to probe the limits of models that researchers use to predict the behavior of materials and the progress of chemical reactions. More

    • 125 Shares169 Views

      in Data Management & Statistics

      Last-mile routing research challenge awards $175,000 to three winning teams

      by

      Routing is one of the most studied problems in operations research; even small improvements in routing efficiency can save companies money and result in energy savings and reduced environmental impacts. Now, three teams of researchers from universities around the world have received prize money totaling $175,000 for their innovative route optimization models.

      The three teams were the winners of the Amazon Last-Mile Routing Research Challenge, through which the MIT Center for Transportation & Logistics (MIT CTL) and Amazon engaged with a global community of researchers across a range of disciplines, from computer science to business operations to supply chain management, challenging them to build data-driven route optimization models leveraging massive historical route execution data.

      First announced in February, the research challenge attracted more than 2,000 participants from around the world. Two hundred twenty-nine researcher teams formed during the spring to independently develop solutions that incorporated driver know-how into route optimization models with the intent that they would outperform traditional optimization approaches. Out of the 48 teams whose models qualified for the final round of the challenge, three teams’ work stood out above the rest. Amazon provided real operational training data for the models and evaluated submissions, with technical support from MIT CTL scientists.

      In real life, drivers frequently deviate from planned and mathematically optimized route sequences. Drivers carry information about which roads are hard to navigate when traffic is bad, when and where they can easily find parking, which stops can be conveniently served together, and many other factors that existing optimization models simply don’t capture.

      Each model addressed the challenge data in a unique way. The methodological approaches chosen by the participants frequently combined traditional exact and heuristic optimization approaches with nontraditional machine learning methods. On the machine learning side, the most commonly adopted methods were different variants of artificial neural networks, as well as inverse reinforcement learning approaches.

      There were 45 submissions that reached the finalist phase, with team members hailing from 29 countries. Entrants spanned all levels of higher education from final-year undergraduate students to retired faculty. Entries were assessed in a double-blind review process so that the judges would not know what team was attached to each entry.

      The third-place prize of $25,000 was awarded to Okan Arslan and Rasit Abay. Okan is a professor at HEC Montréal, and Rasit is a doctoral student at the University of New South Wales in Australia. The runner-up prize at $50,000 was awarded to MIT’s own Xiaotong Guo, Qingyi Wang, and Baichuan Mo, all doctoral students. The top prize of $100,000 was awarded to Professor William Cook of the University of Waterloo in Canada, Professor Stephan Held of the University of Bonn in Germany, and Professor Emeritus Keld Helsgaun of Roskilde University in Denmark. Congratulations to all winners and contestants were held via webinar on July 30.

      Top-performing teams may be interviewed by Amazon for research roles in the company’s Last Mile organization. MIT CTL will publish and promote short technical papers written by all finalists and might invite top-performing teams to present at MIT. Further, a team led by Matthias Winkenbach, director of the MIT Megacity Logistics Lab, will guest-edit a special issue of Transportation Science, one of the most renowned academic journals in this field, featuring academic papers on topics related to the problem tackled by the research challenge. 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