Computers Math

For decades, there has been a trend in microelectronics towards ever smaller and more compact transistors. 2D materials such as graphene are seen as a beacon of hope here: they are the thinnest material layers that can possibly exist, consisting of only one or a few atomic layers. Nevertheless, they can conduct electrical currents — conventional silicon technology, on the other hand, no longer works properly if the layers become too thin.
However, such materials are not used in a vacuum; they have to be combined with suitable insulators — in order to seal them off from unwanted environmental influences, and also in order to control the flow of current via the so-called field effect. Until now, hexagonal boron nitride (hBN) has frequently been used for this purpose as it forms an excellent environment for 2D materials. However, studies conducted by TU Wien, in cooperation with ETH Zurich, the Russian Ioffe Institute and researchers from Saudi Arabia and Japan, now show that, contrary to previous assumptions, thin hBN layers are not suitable as insulators for future miniaturised field-effect transistors, as exorbitant leakage currents occur. So if 2D materials are really to revolutionise the semiconductor industry, one has to start looking for other insulator materials. The study has now been published in the scientific journal “Nature Electronics.”
The supposedly perfect insulator material
“At first glance, hexagonal boron nitride fits graphene and two-dimensional materials better than any other insulator,” says Theresia Knobloch, first author of the study, who is currently working on her dissertation in Tibor Grasser’s team at the Institute of Microelectronics at TU Wien. “Just like the 2D semiconducting materials, hBN consists of individual atomic layers that are only weakly bonded to each other.”
As a result, hBN can easily be used to make atomically smooth surfaces that do not interfere with the transport of electrons through 2D materials. “You might therefore think that hBN is the perfect material — both as a substrate on which to place thin-film semiconductors and also as a gate insulator needed to build field-effect transistors,” says Tibor Grasser.
Small leakage currents with big effects
A transistor can be compared to a water tap — only instead of a stream of water, electric current is switched on and off. As with a water tap, it is very important for a transistor that nothing leaks out of the valve itself.
This is exactly what the gate insulator is responsible for in the transistor: It isolates the controlling electrode, via which the current flow is switched on and off, from the semiconducting channel itself, through which the current then flows. A modern microprocessor contains about 50 billion transistors — so even a small loss of current at the gates can play an enormous role, because it significantly increases the total energy consumption.
In this study, the research team investigated the leakage currents that flow through thin hBN layers, both experimentally and using theoretical calculations. They found that some of the properties that make hBN such a suitable substrate also significantly increase the leakage currents through hBN. Boron nitride has a small dielectric constant, which means that the material interacts only weakly with electric fields. In consequence, the hBN layers used in miniaturised transistors must only be a few atomic layers thick so that the gate’s electric field can sufficiently control the channel. At the same time, however, the leakage currents become too large in this case, as they increase exponentially when reducing the layer thickness.
The search for insulators
“Our results show that hBN is not suitable as a gate insulator for miniaturised transistors based on 2D materials,” says Tibor Grasser. “This finding is an important guide for future studies, but it is only the beginning of the search for suitable insulators for the smallest transistors. Currently, no known material system can meet all the requirements, but it is only a matter of time and resources until a suitable material system is found.”
“The problem is complex, but this makes it all the more important that many scientists devote themselves to the search for a solution, because our society will need small, fast and, above all, energy-efficient computer chips in the future,” Theresia Knobloch is convinced. More

Researchers at Charité — Universitätsmedizin Berlin and TU Berlin as well as the University of Oslo have developed a new tissue-section analysis system for diagnosing breast cancer based on artificial intelligence (AI). Two further developments make this system unique: For the first time, morphological, molecular and histological data are integrated in a single analysis. Secondly, the system provides a clarification of the AI decision process in the form of heatmaps. Pixel by pixel, these heatmaps show which visual information influenced the AI decision process and to what extent, thus enabling doctors to understand and assess the plausibility of the results of the AI analysis. This represents a decisive and essential step forward for the future regular use of AI systems in hospitals. The results of this research have now been published in Nature Machine Intelligence.
Cancer treatment is increasingly concerned with the molecular characterization of tumor tissue samples. Studies are conducted to determine whether and/or how the DNA has changed in the tumor tissue as well as the gene and protein expression in the tissue sample. At the same time, researchers are becoming increasingly aware that cancer progression is closely related to intercellular cross-talk and the interaction of neoplastic cells with the surrounding tissue — including the immune system.
Although microscopic techniques enable biological processes to be studied with high spatial detail, they only permit a limited measurement of molecular markers. These are rather determined using proteins or DNA taken from tissue. As a result, spatial detail is not possible and the relationship between these markers and the microscopic structures is typically unclear. “We know that in the case of breast cancer, the number of immigrated immune cells, known as lymphocytes, in tumor tissue has an influence on the patient’s prognosis. There are also discussions as to whether this number has a predictive value — in other words if it enables us to say how effective a particular therapy is,” says Prof. Dr. Frederick Klauschen of Charité’s Institute of Pathology.
“The problem we have is the following: We have good and reliable molecular data and we have good histological data with high spatial detail. What we don’t have as yet is the decisive link between imaging data and high-dimensional molecular data,” adds Prof. Dr. Klaus-Robert Müller, professor of machine learning at TU Berlin. Both researchers have been working together for a number of years now at the national AI center of excellence the Berlin Institute for the Foundations of Learning and Data (BIFOLD) located at TU Berlin.
It is precisely this symbiosis which the newly published approach makes possible. “Our system facilitates the detection of pathological alterations in microscopic images. Parallel to this, we are able to provide precise heatmap visualizations showing which pixel in the microscopic image contributed to the diagnostic algorithm and to what extent,” explains Prof. Müller. The research team has also succeeded in significantly further developing this process: “Our analysis system has been trained using machine learning processes so that it can also predict various molecular characteristics, including the condition of the DNA, the gene expression as well as the protein expression in specific areas of the tissue, on the basis of the histological images.
Next on the agenda are certification and further clinical validations — including tests in tumor routine diagnostics. However, Prof. Klauschen is already convinced of the value of the research: “The methods we have developed will make it possible in the future to make histopathological tumor diagnostics more precise, more standardized and qualitatively better.”

Story Source:
Materials provided by Charité – Universitätsmedizin Berlin. Note: Content may be edited for style and length. More

Smart speakers, such as Amazon Echo and Google Home, have proven adept at monitoring certain health care issues at home. For example, researchers at the University of Washington have shown that these devices can detect cardiac arrests or monitor babies breathing.
But what about tracking something even smaller: the minute motion of individual heartbeats in a person sitting in front of a smart speaker?
UW researchers have developed a new skill for a smart speaker that for the first time monitors both regular and irregular heartbeats without physical contact. The system sends inaudible sounds from the speaker out into a room and, based on the way the sounds are reflected back to the speaker, it can identify and monitor individual heartbeats. Because the heartbeat is such a tiny motion on the chest surface, the team’s system uses machine learning to help the smart speaker locate signals from both regular and irregular heartbeats.
When the researchers tested this system on healthy participants and hospitalized cardiac patients, the smart speaker detected heartbeats that closely matched the beats detected by standard heartbeat monitors. The team published these findings March 9 in Communications Biology.
“Regular heartbeats are easy enough to detect even if the signal is small, because you can look for a periodic pattern in the data,” said co-senior author Shyam Gollakota, a UW associate professor in the Paul G. Allen School of Computer Science & Engineering. “But irregular heartbeats are really challenging because there is no such pattern. I wasn’t sure that it would be possible to detect them, so I was pleasantly surprised that our algorithms could identify irregular heartbeats during tests with cardiac patients.”
While many people are familiar with the concept of a heart rate, doctors are more interested in the assessment of heart rhythm. Heart rate is the average of heartbeats over time, whereas a heart rhythm describes the pattern of heartbeats.

advertisement

For example, if a person has a heart rate of 60 beats per minute, they could have a regular heart rhythm — one beat every second — or an irregular heart rhythm — beats are randomly scattered across that minute but they still average out to 60 beats per minute.
“Heart rhythm disorders are actually more common than some other well-known heart conditions. Cardiac arrhythmias can cause major morbidities such as strokes, but can be highly unpredictable in occurrence, and thus difficult to diagnose,” said co-senior author Dr. Arun Sridhar, assistant professor of cardiology at the UW School of Medicine. “Availability of a low-cost test that can be performed frequently and at the convenience of home can be a game-changer for certain patients in terms of early diagnosis and management.”
The key to assessing heart rhythm lies in identifying the individual heartbeats. For this system, the search for heartbeats begins when a person sits within 1 to 2 feet in front of the smart speaker. Then the system plays an inaudible continuous sound, which bounces off the person and then returns to the speaker. Based on how the returned sound has changed, the system can isolate movements on the person — including the rise and fall of their chest as they breathe.
“The motion from someone’s breathing is orders of magnitude larger on the chest wall than the motion from heartbeats, so that poses a pretty big challenge,” said lead author Anran Wang, a doctoral student in the Allen School. “And the breathing signal is not regular so it’s hard to simply filter it out. Using the fact that smart speakers have multiple microphones, we designed a new beam-forming algorithm to help the speakers find heartbeats.”
The team designed what’s called a self-supervised machine learning algorithm, which learns on the fly instead of from a training set. This algorithm combines signals from all of the smart speaker’s multiple microphones to identify the elusive heartbeat signal.

advertisement

“This is similar to how Alexa can always find my voice even if I’m playing a video or if there are multiple people talking in the room,” Gollakota said. “When I say, ‘Hey, Alexa,’ the microphones are working together to find me in the room and listen to what I say next. That’s basically what’s happening here but with the heartbeat.”
The heartbeat signals that the smart speaker detects don’t look like the typical peaks that are commonly associated with traditional heartbeat monitors. The researchers used a second algorithm to segment the signal into individual heartbeats so that the system could extract what is known as the inter-beat interval, or the amount of time between two heartbeats.
“With this method, we are not getting the electric signal of the heart contracting. Instead we’re seeing the vibrations on the skin when the heart beats,” Wang said.
The researchers tested a prototype smart speaker running this system on two groups: 26 healthy participants and 24 hospitalized patients with a diversity of cardiac conditions, including atrial fibrillation and heart failure. The team compared the smart speaker’s inter-beat interval with one from a standard heartbeat monitor. Of the nearly 12,300 heartbeats measured for the healthy participants, the smart speaker’s median inter-beat interval was within 28 milliseconds of the standard monitor. The smart speaker performed almost as well with cardiac patients: of the more than 5,600 heartbeats measured, the median inter-beat interval was within 30 milliseconds of the standard.
Currently this system is set up for spot checks: If a person is concerned about their heart rhythm, they can sit in front of a smart speaker to get a reading. But the research team hopes that future versions could continuously monitor heartbeats while people are asleep, something that could help doctors diagnose conditions such as sleep apnea.
“If you have a device like this, you can monitor a patient on an extended basis and define patterns that are individualized for the patient. For example, we can figure out when arrhythmias are happening for each specific patient and then develop corresponding care plans that are tailored for when the patients actually need them,” Sridhar said. “This is the future of cardiology. And the beauty of using these kinds of devices is that they are already in people’s homes.”
This research was funded by the National Science Foundation. More

Anxiety levels in the United States are rising sharply and have especially intensified in younger populations. According to the Anxiety and Depression Association of America, anxiety disorders affect 31.9 percent of children ages 13 to 18 years old. Because of the COVID-19 pandemic, children and adolescents have experienced unprecedented interruptions to their daily lives and it is expected that these disruptions may precipitate mental illness, including anxiety, depression, and/or stress related symptoms.
Traditional anxiety and depression treatments include cognitive behavioral therapy and psychiatric medications, which are somewhat successful in alleviating symptoms in adults. However, they have yielded some mixed results in children. Therefore, discovering appropriate means for reducing childhood anxiety and depression that are both affordable and accessible is paramount.
Using a computerized and completely remote training program, researchers from Florida Atlantic University’s Charles E. Schmidt College of Science have found a way to alleviate negative emotions in preadolescent children. They examined the relationship between anxiety, inhibitory control, and resting-state electroencephalography (EEG) in a critical age-range for social and emotional development (ages 8 to 12 years old). Inhibitory control is the ability to willfully withhold or suppress a thought, action or feeling. Without it, people would act purely on impulses or on old habits of action and thought.
Results of the study published in the journal, Applied Neuropsychology: Child, reveal that computerized inhibitory training helps to mitigate negative emotions in preadolescent children. EEG results also provide evidence of frontal alpha asymmetry shifting to the left after children completed an emotional version of the training. At the baseline time point, there was further indication to support the link between inhibitory control dysfunction and anxiety/depression. Decreased inhibitory control performance predicted higher levels of anxiety and depression, signifying that inhibitory impairments could be a risk factor for the development of these conditions in children.
Prior research has focused on adults and has only used self-report measures to operationalize anxiety and depressive symptoms. This novel study expands upon research investigating cognitive and neurological mechanisms involved in childhood anxiety and depression. In addition, it includes an objective outcome measure (resting-state EEG) to enable more succinct conclusions about training efficacy.
“In the current social climate of the world, internalizing conditions like anxiety and depression are becoming increasingly common in children and adolescents. Meanwhile, the availability and accessibility of computer and tablet technology also has rapidly increased,” said Nathaniel Shanok, lead author and a recent Ph.D. graduate of FAU’s Department of Psychology, who received an award from the American Psychological Society in 2020 for this research. “Providing computerized cognitive training programs to children can be a highly beneficial use of this technology for improving not only academic performance, but as seen in our study, psychological and emotional functioning during a challenging time period of development.”
Participants in the study were assigned to four weeks of either an emotional inhibitory control training program, a neutral inhibitory control training program, or a waitlisted control, and were tested using cognitive, emotional and EEG measures. Researchers evaluated the effects of the four-week, 16-session computerized inhibitory control training program using three tasks (go/no-go, flanker, and Stroop). The training program utilized for the study is gFocus from IQ Mindware and was created by Mark Ashton Smith, Ph.D.
Researchers found that inhibitory control accuracy was significantly and negatively related to anxiety as well as depression. Emotional and neutral training conditions led to significant reductions in anxiety, depression, and negative affect relative to the waitlist group, with the emotional training condition showing the largest reductions in anxiety and negative affect. These two conditions showed comparable improvements in inhibitory control accuracy relative to the waitlist, with greater increases observed in the neutral training condition.
“Given the predominately adverse influence of anxiety on social, psychological and cognitive functioning; early prevention, management and quality treatment plans are critical research areas to explore,” said Nancy Aaron Jones, Ph.D., co-author, an associate professor, and director of the FAU WAVES Emotion Laboratory in the Department of Psychology, Charles E. Schmidt College of Science, and a member of the FAU Brain Institute. “Advancements in technology have made it possible to train certain cognitive abilities using child-friendly applications or games that can be easily accessed from a home computer. Devising computerized training programs, which target underlying cognitive characteristics related to anxiety is a promising method for attenuating symptoms and risk in children.”
Anxiety involves strong cognitive and emotional influences, which are both explicit (obsessive thought processes and ruminations) as well as implicit (negative processing bias and reduced cognitive control). Inhibitory control impairment is believed to be a cognitive mediator of anxiety through both explicit and implicit mechanisms. The explicit theory of generalized anxiety disorder explains that the manifestation of anxiety occurs in part due to the inability of individuals to effectively recognize unrealistic or overly critical thinking patterns (thoughts) and suppress them. More

The perils of machine learning — using computers to identify and analyze data patterns, such as in facial recognition software — have made headlines lately. Yet the technology also holds promise to help enforce federal regulations, including those related to the environment, in a fair, transparent way, according to a new study by Stanford researchers.
The analysis, published this week in the proceedings of the Association of Computing Machinery Conference on Fairness, Accountability and Transparency(link is external), evaluates machine learning techniques designed to support a U.S. Environmental Protection Agency (EPA) initiative to reduce severe violations of the Clean Water Act. It reveals how two key elements of so-called algorithmic design influence which communities are targeted for compliance efforts and, consequently, who bears the burden of pollution violations. The analysis — funded through the Stanford Woods Institute for the Environment’s Realizing Environmental Innovation Program — is timely given recent executive actions(link is external) calling for renewed focus on environmental justice.
“Machine learning is being used to help manage an overwhelming number of things that federal agencies are tasked to do — as a way to help increase efficiency,” said study co-principal investigator Daniel Ho, the William Benjamin Scott and Luna M. Scott Professor of Law at Stanford Law School. “Yet what we also show is that simply designing a machine learning-based system can have an additional benefit.”
Pervasive noncompliance
The Clean Water Act aims to limit pollution from entities that discharge directly into waterways, but in any given year, nearly 30 percent of such facilities self-report persistent or severe violations of their permits. In an effort to halve this type of noncompliance by 2022, EPA has been exploring the use of machine learning to target compliance resources.
To test this approach, EPA reached out to the academic community. Among its chosen partners: Stanford’s Regulation, Evaluation and Governance Lab (RegLab), an interdisciplinary team of legal experts, data scientists, social scientists and engineers that Ho heads. The group has done ongoing work with federal and state agencies to aid environmental compliance.

advertisement

In the new study, RegLab researchers examined how permits with similar functions, such as wastewater treatment plants, were classified by each state in ways that would affect their inclusion in the EPA national compliance initiative. Using machine learning models, they also sifted through hundreds of millions of observations — an impossible task with conventional approaches — from EPA databases on historical discharge volumes, compliance history and permit-level variables to predict the likelihood of future severe violations and the amount of pollution each facility would likely generate. They then evaluated demographic data, such as household income and minority population, for the areas where each model indicated the riskiest facilities were located.
Devil in the details
The team’s algorithmic process helped surface two key ways that the design of the EPA compliance initiative could influence who receives resources. These differences centered on which types of permits were included or excluded, as well as how the goal itself was articulated.
In the process of figuring out how to achieve the compliance goal, the researchers first had to translate the overall objective into a series of concrete instructions — an algorithm — needed to fulfill it. As they were assessing which facilities to run predictions on, they noticed an important embedded decision. While the EPA initiative expands covered permits by at least sevenfold relative to prior efforts, it limits its scope to “individual permits,” which cover a specific discharging entity, such as a single wastewater treatment plant. Left out are “general permits,” intended to cover multiple dischargers engaged in similar activities and with similar types of effluent. A related complication: Most permitting and monitoring authority is vested in state environmental agencies. As a result, functionally similar facilities may be included or excluded from the federal initiative based on how states implement their pollution permitting process.
“The impact of this environmental federalism makes partnership with states critical to achieving these larger goals in an equitable way,” said co-author Reid Whitaker, a RegLab affiliate and 2020 graduate of Stanford Law School now pursuing a PhD in the Jurisprudence and Social Policy Program at the University of California, Berkeley.
Second, the current EPA initiative focuses on reducing rates of noncompliance. While there are good reasons for this policy goal, the researchers’ algorithmic design process made clear that favoring this over pollution discharges that exceed the permitted limit would have a powerful unintended effect. Namely, it would shift enforcement resources away from the most severe violators, which are more likely to be in densely populated minority communities, and toward smaller facilities in more rural, predominantly white communities, according to the researchers.
“Breaking down the big idea of the compliance initiative into smaller chunks that a computer could understand forced a conversation about making implicit decisions explicit,” said study lead author Elinor Benami, a faculty affiliate at the RegLab and assistant professor of agricultural and applied economics at Virginia Tech. “Careful algorithmic design can help regulators transparently identify how objectives translate to implementation while using these techniques to address persistent capacity constraints.” More

The public doesn’t need to know how Artificial Intelligence works to trust it. They just need to know that someone with the necessary skillset is examining AI and has the authority to mete out sanctions if it causes or is likely to cause harm.
Dr Bran Knowles, a senior lecturer in data science at Lancaster University, says: “I’m certain that the public are incapable of determining the trustworthiness of individual AIs… but we don’t need them to do this. It’s not their responsibility to keep AI honest.”
Dr Knowles presents (March 8) a research paper ‘The Sanction of Authority: Promoting Public Trust in AI’ at the ACM Conference on Fairness, Accountability and Transparency (ACM FAccT).
The paper is co-authored by John T. Richards, of IBM’s T.J. Watson Research Center, Yorktown Heights, New York.
The general public are, the paper notes, often distrustful of AI, which stems both from the way AI has been portrayed over the years and from a growing awareness that there is little meaningful oversight of it.
The authors argue that greater transparency and more accessible explanations of how AI systems work, perceived to be a means of increasing trust, do not address the public’s concerns.

advertisement

A ‘regulatory ecosystem’, they say, is the only way that AI will be meaningfully accountable to the public, earning their trust.
“The public do not routinely concern themselves with the trustworthiness of food, aviation, and pharmaceuticals because they trust there is a system which regulates these things and punishes any breach of safety protocols,” says Dr Richards.
And, adds Dr Knowles: “Rather than asking that the public gain skills to make informed decisions about which AIs are worthy of their trust, the public needs the same guarantees that any AI they might encounter is not going to cause them harm.”
She stresses the critical role of AI documentation in enabling this trustworthy regulatory ecosystem. As an example, the paper discusses work by IBM on AI Factsheets, documentation designed to capture key facts regarding an AI’s development and testing.
But, while such documentation can provide information needed by internal auditors and external regulators to assess compliance with emerging frameworks for trustworthy AI, Dr Knowles cautions against relying on it to directly foster public trust.
“If we fail to recognise that the burden to oversee trustworthiness of AI must lie with highly skilled regulators, then there’s a good chance that the future of AI documentation is yet another terms and conditions-style consent mechanism — something no one really reads or understands,” she says.
The paper calls for AI documentation to be properly understood as a means to empower specialists to assess trustworthiness.
“AI has material consequences in our world which affect real people; and we need genuine accountability to ensure that the AI that pervades our world is helping to make that world better,” says Dr Knowles.
ACM FAccT is a computer science conference that brings together researchers and practitioners interested in fairness, accountability, and transparency in socio-technical systems.

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

In a perfect world, what you see is what you get. If this were the case, the job of artificial intelligence systems would be refreshingly straightforward.
Take collision avoidance systems in self-driving cars. If visual input to on-board cameras could be trusted entirely, an AI system could directly map that input to an appropriate action — steer right, steer left, or continue straight — to avoid hitting a pedestrian that its cameras see in the road.
But what if there’s a glitch in the cameras that slightly shifts an image by a few pixels? If the car blindly trusted so-called “adversarial inputs,” it might take unnecessary and potentially dangerous action.
A new deep-learning algorithm developed by MIT researchers is designed to help machines navigate in the real, imperfect world, by building a healthy “skepticism” of the measurements and inputs they receive.
The team combined a reinforcement-learning algorithm with a deep neural network, both used separately to train computers in playing video games like Go and chess, to build an approach they call CARRL, for Certified Adversarial Robustness for Deep Reinforcement Learning.
The researchers tested the approach in several scenarios, including a simulated collision-avoidance test and the video game Pong, and found that CARRL performed better — avoiding collisions and winning more Pong games — over standard machine-learning techniques, even in the face of uncertain, adversarial inputs.

advertisement

“You often think of an adversary being someone who’s hacking your computer, but it could also just be that your sensors are not great, or your measurements aren’t perfect, which is often the case,” says Michael Everett, a postdoc in MIT’s Department of Aeronautics and Astronautics (AeroAstro). “Our approach helps to account for that imperfection and make a safe decision. In any safety-critical domain, this is an important approach to be thinking about.”
Everett is the lead author of a study outlining the new approach, which appears in IEEE’s Transactions on Neural Networks and Learning Systems. The study originated from MIT PhD student Björn Lütjens’ master’s thesis and was advised by MIT AeroAstro Professor Jonathan How.
Possible realities
To make AI systems robust against adversarial inputs, researchers have tried implementing defenses for supervised learning. Traditionally, a neural network is trained to associate specific labels or actions with given inputs. For instance, a neural network that is fed thousands of images labeled as cats, along with images labeled as houses and hot dogs, should correctly label a new image as a cat.
In robust AI systems, the same supervised-learning techniques could be tested with many slightly altered versions of the image. If the network lands on the same label — cat — for every image, there’s a good chance that, altered or not, the image is indeed of a cat, and the network is robust to any adversarial influence.

advertisement

But running through every possible image alteration is computationally exhaustive and difficult to apply successfully to time-sensitive tasks such as collision avoidance. Furthermore, existing methods also don’t identify what label to use, or what action to take, if the network is less robust and labels some altered cat images as a house or a hotdog.
“In order to use neural networks in safety-critical scenarios, we had to find out how to take real-time decisions based on worst-case assumptions on these possible realities,” Lütjens says.
The best reward
The team instead looked to build on reinforcement learning, another form of machine learning that does not require associating labeled inputs with outputs, but rather aims to reinforce certain actions in response to certain inputs, based on a resulting reward. This approach is typically used to train computers to play and win games such as chess and Go.
Reinforcement learning has mostly been applied to situations where inputs are assumed to be true. Everett and his colleagues say they are the first to bring “certifiable robustness” to uncertain, adversarial inputs in reinforcement learning.
Their approach, CARRL, uses an existing deep-reinforcement-learning algorithm to train a deep Q-network, or DQN — a neural network with multiple layers that ultimately associates an input with a Q value, or level of reward.
The approach takes an input, such as an image with a single dot, and considers an adversarial influence, or a region around the dot where it actually might be instead. Every possible position of the dot within this region is fed through a DQN to find an associated action that would result in the most optimal worst-case reward, based on a technique developed by recent MIT graduate student Tsui-Wei “Lily” Weng PhD ’20.
An adversarial world
In tests with the video game Pong, in which two players operate paddles on either side of a screen to pass a ball back and forth, the researchers introduced an “adversary” that pulled the ball slightly further down than it actually was. They found that CARRL won more games than standard techniques, as the adversary’s influence grew.
“If we know that a measurement shouldn’t be trusted exactly, and the ball could be anywhere within a certain region, then our approach tells the computer that it should put the paddle in the middle of that region, to make sure we hit the ball even in the worst-case deviation,” Everett says.
The method was similarly robust in tests of collision avoidance, where the team simulated a blue and an orange agent attempting to switch positions without colliding. As the team perturbed the orange agent’s observation of the blue agent’s position, CARRL steered the orange agent around the other agent, taking a wider berth as the adversary grew stronger, and the blue agent’s position became more uncertain.
There did come a point when CARRL became too conservative, causing the orange agent to assume the other agent could be anywhere in its vicinity, and in response completely avoid its destination. This extreme conservatism is useful, Everett says, because researchers can then use it as a limit to tune the algorithm’s robustness. For instance, the algorithm might consider a smaller deviation, or region of uncertainty, that would still allow an agent to achieve a high reward and reach its destination.
In addition to overcoming imperfect sensors, Everett says CARRL may be a start to helping robots safely handle unpredictable interactions in the real world.
“People can be adversarial, like getting in front of a robot to block its sensors, or interacting with them, not necessarily with the best intentions,” Everett says. “How can a robot think of all the things people might try to do, and try to avoid them? What sort of adversarial models do we want to defend against? That’s something we’re thinking about how to do.”
This research was supported, in part, by Ford Motor Company as part of the Ford-MIT Alliance. More

Scientists have taken a major step forward in harnessing machine learning to accelerate the design for better batteries: Instead of using it just to speed up scientific analysis by looking for patterns in data, as researchers generally do, they combined it with knowledge gained from experiments and equations guided by physics to discover and explain a process that shortens the lifetimes of fast-charging lithium-ion batteries.
It was the first time this approach, known as “scientific machine learning,” has been applied to battery cycling, said Will Chueh, an associate professor at Stanford University and investigator with the Department of Energy’s SLAC National Accelerator Laboratory who led the study. He said the results overturn long-held assumptions about how lithium-ion batteries charge and discharge and give researchers a new set of rules for engineering longer-lasting batteries.
The research, reported today in Nature Materials, is the latest result from a collaboration between Stanford, SLAC, the Massachusetts Institute of Technology and Toyota Research Institute (TRI). The goal is to bring together foundational research and industry know-how to develop a long-lived electric vehicle battery that can be charged in 10 minutes.
“Battery technology is important for any type of electric powertrain,” said Patrick Herring, senior research scientist for Toyota Research Institute. “By understanding the fundamental reactions that occur within the battery we can extend its life, enable faster charging and ultimately design better battery materials. We look forward to building on this work through future experiments to achieve lower-cost, better-performing batteries.”
A trio of advances
The new study builds on two previous advances where the group used more conventional forms of machine learning to dramatically accelerate both battery testing and the process of winnowing down many possible charging methods to find the ones that work best. While these studies allowed researchers to make much faster progress — reducing the time needed to determine battery lifetimes by 98%, for instance — they didn’t reveal the underlying physics or chemistry that made some batteries last longer than others, as the latest study did.

advertisement

Combining all three approaches could potentially slash the time needed to bring a new battery technology from the lab bench to the consumer by as much as two-thirds, Chueh said.
“In this case, we are teaching the machine how to learn the physics of a new type of failure mechanism that could help us design better and safer fast-charging batteries,” Chueh said. “Fast charging is incredibly stressful and damaging to batteries, and solving this problem is key to expanding the nation’s fleet of electric vehicles as part of the overall strategy for fighting climate change.”
The new combined approach can also be applied to developing the grid-scale battery systems needed for a greater deployment of wind and solar electricity, which will become even more urgent as the nation pursues recently announced Biden Administration goals of eliminating fossil fuels from electric power generation by 2035 and achieving net-zero carbon emissions by 2050.
Zooming in for closeups
The new study zoomed in on battery electrodes, which are made of nano-sized grains glommed together into particles. Lithium ions slosh back and forth between the cathode and anode during charging and discharging, seeping into the particles and flowing back out again. This constant back-and-forth makes particles swell, shrink and crack, gradually decreasing their ability to store charge, and fast charging just makes things worse.

advertisement

To look at this process in more detail, the team observed the behavior of cathode particles made of nickel, manganese and cobalt, a combination known as NMC that’s one of the most widely used materials in electric vehicle batteries. These particles absorb lithium ions when the battery discharges and release them when it charges.
Stanford postdoctoral researchers Stephen Dongmin Kang and Jungjin Park used X-rays from SLAC’s Stanford Synchrotron Radiation Lightsource to get an overall look at particles that were undergoing fast charging. Then they took particles to Lawrence Berkeley National Laboratory’s Advanced Light Source to be examined with scanning X-ray transmission microscopy, which homes in on individual particles.
The data from those experiments, along with information from mathematical models of fast charging and equations that describe the chemistry and physics of the process, were incorporated into scientific machine learning algorithms.
“Rather than having the computer directly figure out the model by simply feeding it data, as we did in the two previous studies, we taught the computer how to choose or learn the right equations, and thus the right physics,” said Kang, who performed the modeling with MIT graduate student Hongbo Zhao, working with chemical engineering professor Martin Bazant.
The rich-get-richer effect
Until now, scientists had assumed that the differences between particles were insignificant and that their ability to store and release ions was limited by how fast lithium could move inside the particles, Kang said. In this way of seeing things, lithium ions flow in and out of all the particles at the same time and at roughly the same speed.
But the new approach revealed that the particles themselves control how fast lithium ions move out of cathode particles when a battery charges, he said. Some particles immediately release a lot of their ions while others release very few or none at all. And the quick-to-release particles go on releasing ions at a faster rate than their neighbors – a positive feedback, or “rich get richer,” effect that had not been identified before.
“We now have a picture — literally a movie — of how lithium moves around inside the battery, and it’s very different than scientists and engineers thought it was,” Kang said. “This uneven charging and discharging puts more stress on the electrodes and decreases their working lifetimes. Understanding this process on a fundamental level is an important step toward solving the fast charging problem.”
The scientists say their new method has potential for improving the cost, storage capacity, durability and other important properties of batteries for a wide range of applications, from electric vehicles to laptops to large-scale storage of renewable energy on the grid.
“It was just two years ago that the 2019 Nobel Prize in chemistry was awarded for the development of rechargeable lithium-ion batteries, which dates back to the 1970s,” Chueh said. “So I am encouraged that there’s still so much to learn about how to make batteries better.”
This research was funded by Toyota Research Institute. The Stanford Synchrotron Radiation Lightsource and Advanced Light Source are DOE Office of Science user facilities, and work there was supported by the DOE Office of Science and the DOE Advanced Battery Materials Research Program. More