More stories

  • in

    New insights into memristive devices by combining incipient ferroelectrics and graphene

    Scientists are working on new materials to create neuromorphic computers, with a design based on the human brain. A crucial component is a memristive device, the resistance of which depends on the history of the device — just like the response of our neurons depends on previous input. Materials scientists from the University of Groningen analysed the behaviour of strontium titanium oxide, a platform material for memristor research and used the 2D material graphene to probe it. On 11 November 2020, the results were published in the journal ACS Applied Materials and Interfaces.
    Computers are giant calculators, full of switches that have a value of either 0 or 1. Using a great many of these binary systems, computers can perform calculations very rapidly. However, in other respects, computers are not very efficient. Our brain uses less energy for recognizing faces or performing other complex tasks than a standard microprocessor. That is because our brain is made up of neurons that can have many values other than 0 and 1 and because the neurons’ output depends on previous input.
    Oxygen vacancies
    To create memristors, switches with a memory of past events, strontium titanium oxide (STO) is often used. This material is a perovskite, whose crystal structure depends on temperature, and can become an incipient ferroelectric at low temperatures. The ferroelectric behaviour is lost above 105 Kelvin. The domains and domain walls that accompany these phase transitions are the subject of active research. Yet, it is still not entirely clear why the material behaves the way it does. ‘It is in a league of its own,’ says Tamalika Banerjee, Professor of Spintronics of Functional Materials at the Zernike Institute for Advanced Materials, University of Groningen.
    The oxygen atoms in the crystal appear to be key to its behaviour. ‘Oxygen vacancies can move through the crystal and these defects are important,’ says Banerjee. ‘Furthermore, domain walls are present in the material and they move when a voltage is applied to it.’ Numerous studies have sought to find out how this happens, but looking inside this material is complicated. However, Banerjee’s team succeeded in using another material that is in a league of its own: graphene, the two-dimensional carbon sheet.
    Conductivity
    ‘The properties of graphene are defined by its purity,’ says Banerjee, ‘whereas the properties of STO arise from imperfections in the crystal structure. We found that combining them leads to new insights and possibilities.’ Much of this work was carried out by Banerjee’s PhD student Si Chen. She placed graphene strips on top of a flake of STO and measured the conductivity at different temperatures by sweeping a gate voltage between positive and negative values. ‘When there is an excess of either electrons or the positive holes, created by the gate voltage, graphene becomes conductive,’ Chen explains. ‘But at the point where there are very small amounts of electrons and holes, the Dirac point, conductivity is limited.’
    In normal circumstances, the minimum conductivity position does not change with the sweeping direction of the gate voltage. However, in the graphene strips on top of STO, there is a large separation between the minimum conductivity positions for the forward sweep and the backward sweep. The effect is very clear at 4 Kelvin, but less pronounced at 105 Kelvin or at 150 Kelvin. Analysis of the results, along with theoretical studies carried out at Uppsala University, shows that oxygen vacancies near the surface of the STO are responsible.
    Memory
    Banerjee: ‘The phase transitions below 105 Kelvin stretch the crystal structure, creating dipoles. We show that oxygen vacancies accumulate at the domain walls and that these walls offer the channel for the movement of oxygen vacancies. These channels are responsible for memristive behaviour in STO.’ Accumulation of oxygen vacancy channels in the crystal structure of STO explains the shift in the position of the minimum conductivity.
    Chen also carried out another experiment: ‘We kept the STO gate voltage at -80 V and measured the resistance in the graphene for almost half an hour. In this period, we observed a change in resistance, indicating a shift from hole to electron conductivity.’ This effect is primarily caused by the accumulation of oxygen vacancies at the STO surface.
    All in all, the experiments show that the properties of the combined STO/graphene material change through the movement of both electrons and ions, each at different time scales. Banerjee: ‘By harvesting one or the other, we can use the different response times to create memristive effects, which can be compared to short-term or long-term memory effects.’ The study creates new insights into the behaviour of STO memristors. ‘And the combination with graphene opens up a new path to memristive heterostructures combining ferroelectric materials and 2D materials.’

    Story Source:
    Materials provided by University of Groningen. Note: Content may be edited for style and length. More

  • in

    Improving quantum dot interactions, one layer at a time

    Osaka City University scientists and colleagues in Japan have found a way to control an interaction between quantum dots that could greatly improve charge transport, leading to more efficient solar cells. Their findings were published in the journal Nature Communications.
    Nanomaterials engineer DaeGwi Kim led a team of scientists at Osaka City University, RIKEN Center for Emergent Matter Science and Kyoto University to investigate ways to control a property called quantum resonance in layered structures of quantum dots called superlattices.
    “Our simple method for fine-tuning quantum resonance is an important contribution to both optical materials and nanoscale material processing,” says Kim.
    Quantum dots are nanometer-sized semiconductor particles with interesting optical and electronic properties. When light is shone on them, for example, they emit strong light at room temperature, a property called photoluminescence. When quantum dots are close enough to each other, their electronic states are coupled, a phenomenon called quantum resonance. This greatly improves their ability to transport electrons between them. Scientists have been wanting to manufacture devices using this interaction, including solar cells, display technologies, and thermoelectric devices.
    However, they have so far found it difficult to control the distances between quantum dots in 1D, 2D and 3D structures. Current fabrication processes use long ligands to hold quantum dots together, which hinders their interactions.
    Kim and his colleagues found they could detect and control quantum resonance by using cadmium telluride quantum dots connected with short N-acetyl-L-cysteine ligands. They controlled the distance between quantum dot layers by placing a spacer layer between them made of oppositely charged polyelectrolytes. Quantum resonance is detected between stacked dots when the spacer layer is thinner than two nanometers. The scientists also controlled the distance between quantum dots in a single layer, and thus quantum resonance, by changing the concentration of quantum dots used in the layering process.
    The team next plans to study the optical properties, especially photoluminescence, of quantum dot superlattices made using their layer-by-layer approach. “This is extremely important for realizing new optical electronic devices made with quantum dot superlattices,” says Kim.
    Kim adds that their fabrication method can be used with other types of water-soluble quantum dots and nanoparticles. “Combining different types of semiconductor quantum dots, or combining semiconductor quantum dots with other nanoparticles, will expand the possibilities of new material design,” says Kim.

    Story Source:
    Materials provided by Osaka City University. Note: Content may be edited for style and length. More

  • in

    Hidden 15th-century text on medieval manuscripts

    Rochester Institute of Technology students discovered lost text on 15th-century manuscript leaves using an imaging system they developed as freshmen. By using ultraviolet-fluorescence imaging, the students revealed that a manuscript leaf held in RIT’s Cary Graphic Arts Collection was actually a palimpsest, a manuscript on parchment with multiple layers of writing.
    At the time the manuscript was written, making parchment was expensive, so leaves were regularly scraped or erased and re-used. While the erased text is invisible to the naked eye, the chemical signature of the initial writing can sometimes be detected using other areas of the light spectrum.
    “Using our system, we borrowed several parchments from the Cary Collection here at RIT and when we put one of them under the UV light, it showed this amazing dark French cursive underneath,” said Zoë LaLena, a second-year imaging science student from Fairport, N.Y., who worked on the project. “This was amazing because this document has been in the Cary Collection for about a decade now and no one noticed. And because it’s also from the Ege Collection, in which there’s 30 other known pages from this book, it’s really fascinating that the 29 other pages we know the location of have the potential to also be palimpsests.”
    The imaging system was originally built by 19 students enrolled in the Chester F. Carlson Center for Imaging Science’s Innovative Freshman Experience, a yearlong, project-based course that has the imaging science, motion picture science, and photographic sciences programs combine their talents to solve a problem.
    When RIT switched to remote instruction in March due to the coronavirus outbreak, the students were unable to finish building it, but thanks to a donation from Jeffrey Harris ’75 (photographic science and instrumentation) and Joyce Pratt, three students received funding to continue to work on the project over the summer. Those three students — LaLena; Lisa Enochs, a second-year student double majoring in motion picture science and imaging science from Mississauga, Ontario; and Malcom Zale, a second-year motion picture science student from Milford, Mass. — finished assembling the system in the fall when classes resumed and began analyzing documents from the Cary Collection.
    Steven Galbraith, curator of the Cary Graphic Arts Collection, said he was excited they discovered the manuscript leaf was a palimpsest because similar leaves have been studied extensively by scholars across the country, but never tested with UV light or fully imaged.
    Collector, educator, and historian Otto Ege made leaf collections out of medieval manuscripts that were damaged or incomplete and sold them or distributed them to libraries and special collections across North America, including to the Cary Collection. Galbraith said he’s excited because it means that many other cultural and academic institutions with Ege Collection leaves now may have palimpsests in their collection to study.
    “The students have supplied incredibly important information about at least two of our manuscript leaves here in the collection and in a sense have discovered two texts that we didn’t know were in the collection,” said Galbraith. “Now we have to figure out what those texts are and that’s the power of spectral imaging in cultural institutions. To fully understand our own collections, we need to know the depth of our collections, and imaging science helps reveal all of that to us.”
    The students are interested to see if more manuscript leaves from Ege collections across the country are palimpsests. They imaged another Ege Collection leaf at the Buffalo and Erie County Public Library that turned out to be a palimpsest and are reaching out to other curators across the country. As they begin stitching the lost text back together, paleographers can examine the information they contain.
    The students have been selected to share their results at the 2021 International Congress on Medieval Studies and also plan to present the project at next year’s Imagine RIT: Creativity and Innovation Festival.
    VIDEO: https://www.youtube.com/watch?v=YEieepHPMA0

    Story Source:
    Materials provided by Rochester Institute of Technology. Original written by Luke Auburn. Note: Content may be edited for style and length. More

  • in

    Showing robots how to drive a car…in just a few easy lessons

    Imagine if robots could learn from watching demonstrations: you could show a domestic robot how to do routine chores or set a dinner table. In the workplace, you could train robots like new employees, showing them how to perform many duties. On the road, your self-driving car could learn how to drive safely by watching you drive around your neighborhood.
    Making progress on that vision, USC researchers have designed a system that lets robots autonomously learn complicated tasks from a very small number of demonstrations — even imperfect ones. The paper, titled Learning from Demonstrations Using Signal Temporal Logic, was presented at the Conference on Robot Learning (CoRL), Nov. 18.
    The researchers’ system works by evaluating the quality of each demonstration, so it learns from the mistakes it sees, as well as the successes. While current state-of-art methods need at least 100 demonstrations to nail a specific task, this new method allows robots to learn from only a handful of demonstrations. It also allows robots to learn more intuitively, the way humans learn from each other — you watch someone execute a task, even imperfectly, then try yourself. It doesn’t have to be a “perfect” demonstration for humans to glean knowledge from watching each other.
    “Many machine learning and reinforcement learning systems require large amounts of data data and hundreds of demonstrations — you need a human to demonstrate over and over again, which is not feasible,” said lead author Aniruddh Puranic, a Ph.D. student in computer science at the USC Viterbi School of Engineering.
    “Also, most people don’t have programming knowledge to explicitly state what the robot needs to do, and a human cannot possibly demonstrate everything that a robot needs to know. What if the robot encounters something it hasn’t seen before? This is a key challenge.”
    Learning from demonstrations
    Learning from demonstrations is becoming increasingly popular in obtaining effective robot control policies — which control the robot’s movements — for complex tasks. But it is susceptible to imperfections in demonstrations and also raises safety concerns as robots may learn unsafe or undesirable actions.

    advertisement

    Also, not all demonstrations are equal: some demonstrations are a better indicator of desired behavior than others and the quality of the demonstrations often depends on the expertise of the user providing the demonstrations.
    To address these issues, the researchers integrated “signal temporal logic” or STL to evaluate the quality of demonstrations and automatically rank them to create inherent rewards.
    In other words, even if some parts of the demonstrations do not make any sense based on the logic requirements, using this method, the robot can still learn from the imperfect parts. In a way, the system is coming to its own conclusion about the accuracy or success of a demonstration.
    “Let’s say robots learn from different types of demonstrations — it could be a hands-on demonstration, videos, or simulations — if I do something that is very unsafe, standard approaches will do one of two things: either, they will completely disregard it, or even worse, the robot will learn the wrong thing,” said co-author Stefanos Nikolaidis, a USC Viterbi assistant professor of computer science.
    “In contrast, in a very intelligent way, this work uses some common sense reasoning in the form of logic to understand which parts of the demonstration are good and which parts are not. In essence, this is exactly what also humans do.”
    Take, for example, a driving demonstration where someone skips a stop sign. This would be ranked lower by the system than a demonstration of a good driver. But, if during this demonstration, the driver does something intelligent — for instance, applies their brakes to avoid a crash — the robot will still learn from this smart action.

    advertisement

    Adapting to human preferences
    Signal temporal logic is an expressive mathematical symbolic language that enables robotic reasoning about current and future outcomes. While previous research in this area has used “linear temporal logic,” STL is preferable in this case, said Jyo Deshmukh, a former Toyota engineer and USC Viterbi assistant professor of computer science .
    “When we go into the world of cyber physical systems, like robots and self-driving cars, where time is crucial, linear temporal logic becomes a bit cumbersome, because it reasons about sequences of true/false values for variables, while STL allows reasoning about physical signals.”
    Puranic, who is advised by Deshmukh, came up with the idea after taking a hands-on robotics class with Nikolaidis, who has been working on developing robots to learn from YouTube videos. The trio decided to test it out. All three said they were surprised by the extent of the system’s success and the professors both credit Puranic for his hard work.
    “Compared to a state-of-the-art algorithm, being used extensively in many robotics applications, you see an order of magnitude difference in how many demonstrations are required,” said Nikolaidis.
    The system was tested using a Minecraft-style game simulator, but the researchers said the system could also learn from driving simulators and eventually even videos. Next, the researchers hope to try it out on real robots. They said this approach is well suited for applications where maps are known beforehand but there are dynamic obstacles in the map: robots in household environments, warehouses or even space exploration rovers.
    “If we want robots to be good teammates and help people, first they need to learn and adapt to human preference very efficiently,” said Nikolaidis. “Our method provides that.”
    “I’m excited to integrate this approach into robotic systems to help them efficiently learn from demonstrations, but also effectively help human teammates in a collaborative task.” More

  • in

    Artificial intelligence-based tool may help diagnose opioid addiction earlier

    Researchers have used machine learning, a type of artificial intelligence, to develop a prediction model for the early diagnosis of opioid use disorder. The advance is described in Pharmacology Research & Perspectives.
    The model was generated from information in a commercial claim database from 2006 through 2018 of 10 million medical insurance claims from 550,000 patient records. It relied on data such as demographics, chronic conditions, diagnoses and procedures, and medication prescriptions.
    The tool led to a diagnosis of opioid use disorder that was on average 14.4 months earlier than it was diagnosed clinically.
    “Opioid use disorder has led a very serious epidemic in the U.S. and many other countries, with devastating rates of morbidity and mortality due to missed and delayed diagnoses. The novel ability of our algorithm to identify affected individuals earlier will likely save lives and health care costs,” said senior author Gideon Koren, MD, of Ariel University, in Israel.

    Story Source:
    Materials provided by Wiley. Note: Content may be edited for style and length. More

  • in

    Three reasons why COVID-19 can cause silent hypoxia

    Scientists are still solving the many puzzling aspects of how the novel coronavirus attacks the lungs and other parts of the body. One of the biggest and most life-threatening mysteries is how the virus causes “silent hypoxia,” a condition when oxygen levels in the body are abnormally low, which can irreparably damage vital organs if gone undetected for too long. Now, thanks to computer models and comparisons with real patient data, Boston University biomedical engineers and collaborators from the University of Vermont have begun to crack the mystery.
    Despite experiencing dangerously low levels of oxygen, many people infected with severe cases of COVID-19 sometimes show no symptoms of shortness of breath or difficulty breathing. Hypoxia’s ability to quietly inflict damage is why it’s been coined “silent.” In coronavirus patients, it’s thought that the infection first damages the lungs, rendering parts of them incapable of functioning properly. Those tissues lose oxygen and stop working, no longer infusing the blood stream with oxygen, causing silent hypoxia. But exactly how that domino effect occurs has not been clear until now.
    “We didn’t know [how this] was physiologically possible,” says Bela Suki, a BU College of Engineering professor of biomedical engineering and of materials science and engineering and one of the authors of the study. Some coronavirus patients have experienced what some experts have described as levels of blood oxygen that are “incompatible with life.” Disturbingly, Suki says, many of these patients showed little to no signs of abnormalities when they underwent lung scans.
    To help get to the bottom of what causes silent hypoxia, BU biomedical engineers used computer modeling to test out three different scenarios that help explain how and why the lungs stop providing oxygen to the bloodstream. Their research, which has been published in Nature Communications, reveals that silent hypoxia is likely caused by a combination of biological mechanisms that may occur simultaneously in the lungs of COVID-19 patients, according to biomedical engineer Jacob Herrmann, a research postdoctoral associate in Suki’s lab and the lead author of the new study.
    Normally, the lungs perform the life-sustaining duty of gas exchange, providing oxygen to every cell in the body as we breathe in and ridding us of carbon dioxide each time we exhale. Healthy lungs keep the blood oxygenated at a level between 95 and 100 percent — if it dips below 92 percent, it’s a cause for concern and a doctor might decide to intervene with supplemental oxygen. (Early in the coronavirus pandemic, when clinicians first started sounding the alarm about silent hypoxia, oximeters flew off store shelves as many people, worried that they or their family members might have to recover from milder cases of coronavirus at home, wanted to be able to monitor their blood oxygen levels.)
    The researchers first looked at how COVID-19 impacts the lungs’ ability to regulate where blood is directed. Normally, if areas of the lung aren’t gathering much oxygen due to damage from infection, the blood vessels will constrict in those areas. This is actually a good thing that our lungs have evolved to do, because it forces blood to instead flow through lung tissue replete with oxygen, which is then circulated throughout the rest of the body.
    But according to Herrmann, preliminary clinical data have suggested that the lungs of some COVID-19 patients had lost the ability of restricting blood flow to already damaged tissue, and in contrast, were potentially opening up those blood vessels even more — something that is hard to see or measure on a CT scan.
    Using a computational lung model, Herrmann, Suki, and their team tested that theory, revealing that for blood oxygen levels to drop to the levels observed in COVID-19 patients, blood flow would indeed have to be much higher than normal in areas of the lungs that can no longer gather oxygen — contributing to low levels of oxygen throughout the entire body, they say.
    Next, they looked at how blood clotting may impact blood flow in different regions of the lung. When the lining of blood vessels get inflamed from COVID-19 infection, tiny blood clots too small to be seen on medical scans can form inside the lungs. They found, using computer modeling of the lungs, that this could incite silent hypoxia, but alone it is likely not enough to cause oxygen levels to drop as low as the levels seen in patient data.
    Last, the researchers used their computer model to find out if COVID-19 interferes with the normal ratio of air-to-blood flow that the lungs need to function normally. This type of mismatched air-to-blood flow ratio is something that happens in many respiratory illnesses, such as with asthma patients, Suki says, and it can be a possible contributor to the severe, silent hypoxia that has been observed in COVID-19 patients. Their models suggest that for this to be a cause of silent hypoxia, the mismatch must be happening in parts of the lung that don’t appear injured or abnormal on lung scans.
    Altogether, their findings suggest that a combination of all three factors are likely to be responsible for the severe cases of low oxygen in some COVID-19 patients. By having a better understanding of these underlying mechanisms, and how the combinations could vary from patient to patient, clinicians can make more informed choices about treating patients using measures like ventilation and supplemental oxygen. A number of interventions are currently being studied, including a low-tech intervention called prone positioning that flips patients over onto their stomachs, allowing for the back part of the lungs to pull in more oxygen and evening out the mismatched air-to-blood ratio.
    “Different people respond to this virus so differently,” says Suki. For clinicians, he says it’s critical to understand all the possible reasons why a patient’s blood oxygen might be low, so that they can decide on the proper form of treatment, including medications that could help constrict blood vessels, bust blood clots, or correct a mismatched air-to-blood flow ratio. More

  • in

    Researchers identify features that could make someone a virus super-spreader

    New research from the University of Central Florida has identified physiological features that could make people super-spreaders of viruses such as COVID-19.
    In a study appearing this month in the journal Physics of Fluids, researchers in UCF’s Department of Mechanical and Aerospace Engineering used computer-generated models to numerically simulate sneezes in different types of people and determine associations between people’s physiological features and how far their sneeze droplets travel and linger in the air.
    They found that people’s features, like a stopped-up nose or a full set of teeth, could increase their potential to spread viruses by affecting how far droplets travel when they sneeze.
    According to the U.S. Centers for Disease Control and Prevention, the main way people are infected by the virus that causes COVID-19 is through exposure to respiratory droplets, such as from sneezes and coughs that are carrying infectious virus.
    Knowing more about factors affecting how far these droplets travel can inform efforts to control their spread, says Michael Kinzel, an assistant professor with UCF’s Department of Mechanical Engineering and study co-author.
    “This is the first study that aims to understand the underlying ‘why’ of how far sneezes travel,” Kinzel says. “We show that the human body has influencers, such as a complex duct system associated with the nasal flow that actually disrupts the jet from your mouth and prevents it from dispersing droplets far distances.”
    For instance, when people have a clear nose, such as from blowing it into a tissue, the speed and distance sneeze droplets travel decrease, according to the study.

    advertisement

    This is because a clear nose provides a path in addition to the mouth for the sneeze to exit. But when people’s noses are congested, the area that the sneeze can exit is restricted, thus causing sneeze droplets expelled from the mouth to increase in velocity.
    Similarly, teeth also restrict the sneeze’s exit area and cause droplets to increase in velocity.
    “Teeth create a narrowing effect in the jet that makes it stronger and more turbulent,” Kinzel says. “They actually appear to drive transmission. So, if you see someone without teeth, you can actually expect a weaker jet from the sneeze from them.”
    To perform the study, the researchers used 3D modeling and numerical simulations to recreate four mouth and nose types: a person with teeth and a clear nose; a person with no teeth and a clear nose; a person with no teeth and a congested nose; and a person with teeth and a congested nose.
    When they simulated sneezes in the different models, they found that the spray distance of droplets expelled when a person has a congested nose and a full set of teeth is about 60 percent greater than when they do not.

    advertisement

    The results indicate that when someone keeps their nose clear, such as by blowing it into a tissue, that they could be reducing the distance their germs travel.
    The researchers also simulated three types of saliva: thin, medium and thick.
    They found that thinner saliva resulted in sneezes composed of smaller droplets, which created a spray and stayed in the air longer than medium and thick saliva.
    For instance, three seconds after a sneeze, when thick saliva was reaching the ground and thus diminishing its threat, the thinner saliva was still floating in the air as a potential disease transmitter.
    The work ties back to the researchers’ project to create a COVID-19 cough drop that would give people thicker saliva to reduce the distance droplets from a sneeze or cough would travel, and thus decrease disease-transmission likelihood.
    The findings yield novel insight into variability of exposure distance and indicate how physiological factors affect transmissibility rates, says Kareem Ahmed, an associate professor in UCF’s Department of Mechanical and Aerospace Engineering and study co-author.
    “The results show exposure levels are highly dependent on the fluid dynamics that can vary depending on several human features,” Ahmed says. “Such features may be underlying factors driving superspreading events in the COVID-19 pandemic.”
    The researchers say they hope to move the work toward clinical studies next to compare their simulation findings with those from real people from varied backgrounds.
    Study co-authors were Douglas Fontes, a postdoctoral researcher with the Florida Space Institute and the study’s lead author, and Jonathan Reyes, a postdoctoral researcher in UCF’s Department of Mechanical and Aerospace Engineering.
    Fontes says to advance the findings of the study, the research team wants to investigate the interactions between gas flow, mucus film and tissue structures within the upper respiratory tract during respiratory events.
    “Numerical models and experimental techniques should work side by side to provide accurate predictions of the primary breakup inside the upper respiratory tract during those events,” he says.
    “This research potentially will provide information for more accurate safety measures and solutions to reduce pathogen transmission, giving better conditions to deal with the usual diseases or with pandemics in the future,” he says.
    The work was funded by the National Science Foundation. More

  • in

    A neural network learns when it should not be trusted

    Increasingly, artificial intelligence systems known as deep learning neural networks are used to inform decisions vital to human health and safety, such as in autonomous driving or medical diagnosis. These networks are good at recognizing patterns in large, complex datasets to aid in decision-making. But how do we know they’re correct? Alexander Amini and his colleagues at MIT and Harvard University wanted to find out.
    They’ve developed a quick way for a neural network to crunch data, and output not just a prediction but also the model’s confidence level based on the quality of the available data. The advance might save lives, as deep learning is already being deployed in the real world today. A network’s level of certainty can be the difference between an autonomous vehicle determining that “it’s all clear to proceed through the intersection” and “it’s probably clear, so stop just in case.”
    Current methods of uncertainty estimation for neural networks tend to be computationally expensive and relatively slow for split-second decisions. But Amini’s approach, dubbed “deep evidential regression,” accelerates the process and could lead to safer outcomes. “We need the ability to not only have high-performance models, but also to understand when we cannot trust those models,” says Amini, a PhD student in Professor Daniela Rus’ group at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL).
    “This idea is important and applicable broadly. It can be used to assess products that rely on learned models. By estimating the uncertainty of a learned model, we also learn how much error to expect from the model, and what missing data could improve the model,” says Rus.
    Amini will present the research at next month’s NeurIPS conference, along with Rus, who is the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science, director of CSAIL, and deputy dean of research for the MIT Stephen A. Schwarzman College of Computing; and graduate students Wilko Schwarting of MIT and Ava Soleimany of MIT and Harvard.
    Efficient uncertainty
    After an up-and-down history, deep learning has demonstrated remarkable performance on a variety of tasks, in some cases even surpassing human accuracy. And nowadays, deep learning seems to go wherever computers go. It fuels search engine results, social media feeds, and facial recognition. “We’ve had huge successes using deep learning,” says Amini. “Neural networks are really good at knowing the right answer 99 percent of the time.” But 99 percent won’t cut it when lives are on the line.

    advertisement

    “One thing that has eluded researchers is the ability of these models to know and tell us when they might be wrong,” says Amini. “We really care about that 1 percent of the time, and how we can detect those situations reliably and efficiently.”
    Neural networks can be massive, sometimes brimming with billions of parameters. So it can be a heavy computational lift just to get an answer, let alone a confidence level. Uncertainty analysis in neural networks isn’t new. But previous approaches, stemming from Bayesian deep learning, have relied on running, or sampling, a neural network many times over to understand its confidence. That process takes time and memory, a luxury that might not exist in high-speed traffic.
    The researchers devised a way to estimate uncertainty from only a single run of the neural network. They designed the network with bulked up output, producing not only a decision but also a new probabilistic distribution capturing the evidence in support of that decision. These distributions, termed evidential distributions, directly capture the model’s confidence in its prediction. This includes any uncertainty present in the underlying input data, as well as in the model’s final decision. This distinction can signal whether uncertainty can be reduced by tweaking the neural network itself, or whether the input data are just noisy.
    Confidence check
    To put their approach to the test, the researchers started with a challenging computer vision task. They trained their neural network to analyze a monocular color image and estimate a depth value (i.e. distance from the camera lens) for each pixel. An autonomous vehicle might use similar calculations to estimate its proximity to a pedestrian or to another vehicle, which is no simple task.

    advertisement

    Their network’s performance was on par with previous state-of-the-art models, but it also gained the ability to estimate its own uncertainty. As the researchers had hoped, the network projected high uncertainty for pixels where it predicted the wrong depth. “It was very calibrated to the errors that the network makes, which we believe was one of the most important things in judging the quality of a new uncertainty estimator,” Amini says.
    To stress-test their calibration, the team also showed that the network projected higher uncertainty for “out-of-distribution” data — completely new types of images never encountered during training. After they trained the network on indoor home scenes, they fed it a batch of outdoor driving scenes. The network consistently warned that its responses to the novel outdoor scenes were uncertain. The test highlighted the network’s ability to flag when users should not place full trust in its decisions. In these cases, “if this is a health care application, maybe we don’t trust the diagnosis that the model is giving, and instead seek a second opinion,” says Amini.
    The network even knew when photos had been doctored, potentially hedging against data-manipulation attacks. In another trial, the researchers boosted adversarial noise levels in a batch of images they fed to the network. The effect was subtle — barely perceptible to the human eye — but the network sniffed out those images, tagging its output with high levels of uncertainty. This ability to sound the alarm on falsified data could help detect and deter adversarial attacks, a growing concern in the age of deepfakes.
    Deep evidential regression is “a simple and elegant approach that advances the field of uncertainty estimation, which is important for robotics and other real-world control systems,” says Raia Hadsell, an artificial intelligence researcher at DeepMind who was not involved with the work. “This is done in a novel way that avoids some of the messy aspects of other approaches — e.g. sampling or ensembles — which makes it not only elegant but also computationally more efficient — a winning combination.”
    Deep evidential regression could enhance safety in AI-assisted decision making. “We’re starting to see a lot more of these [neural network] models trickle out of the research lab and into the real world, into situations that are touching humans with potentially life-threatening consequences,” says Amini. “Any user of the method, whether it’s a doctor or a person in the passenger seat of a vehicle, needs to be aware of any risk or uncertainty associated with that decision.” He envisions the system not only quickly flagging uncertainty, but also using it to make more conservative decision making in risky scenarios like an autonomous vehicle approaching an intersection.
    “Any field that is going to have deployable machine learning ultimately needs to have reliable uncertainty awareness,” he says.
    This work was supported, in part, by the National Science Foundation and Toyota Research Institute through the Toyota-CSAIL Joint Research Center. More