Step by step

Berkeley researchers may be one step closer to making robot dogs our new best friends. Using advances in machine learning, two separate teams have developed cutting-edge approaches to shorten in-the-field training times for quadruped robots, getting them to walk — and even roll over — in record time.

In a first for the robotics field, a team led by Sergey Levine, associate professor of electrical engineering and computer sciences, demonstrated a robot learning to walk without prior training from models and simulations in just 20 minutes. The demonstration marks a significant advancement, as this robot relied solely on trial and error in the field to master the movements necessary to walk and adapt to different settings.

“Our work shows that training robots in the real world is more feasible than previously thought, and we hope, as a result, to empower other researchers to start tackling more real-world problems,” said Laura Smith, a Ph.D. student in Levine’s lab and one of the lead authors of the paper posted on arXiv.

In past studies, robots of comparable complexity required several hours to weeks of data input to learn to walk using reinforcement learning (RL). Often, they also were trained in controlled lab settings, where they learned to walk on relatively simple terrain and received precise feedback about their performance.

Reprogrammable materials selectively self-assemble

With just a random disturbance that energizes the cubes, they selectively self-assemble into a larger block. 

Credit: MIT CSAIL

While automated manufacturing is ubiquitous today, it was once a nascent field birthed by inventors such as Oliver Evans, who is credited with creating the first fully automated industrial process, in flour mill he built and gradually automated in the late 1700s. The processes for creating automated structures or machines are still very top-down, requiring humans, factories, or robots to do the assembling and making.

However, the way nature does assembly is ubiquitously bottom-up; animals and plants are self-assembled at a cellular level, relying on proteins to self-fold into target geometries that encode all the different functions that keep us ticking. For a more bio-inspired, bottom-up approach to assembly, then, human-architected materials need to do better on their own. Making them scalable, selective, and reprogrammable in a way that could mimic nature’s versatility means some teething problems, though.

Now, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have attempted to get over these growing pains with a new method: introducing magnetically reprogrammable materials that they coat different parts with — like robotic cubes — to let them self-assemble. Key to their process is a way to make these magnetic programs highly selective about what they connect with, enabling robust self-assembly into specific shapes and chosen configurations.

A Machine Learning-Based Solution Could Help Firefighters Circumvent Deadly Backdrafts

NIST researchers conducted hundreds of fire experiments to find out what conditions make a room ripe for backdraft and fed the data to a machine learning algorithm. The result was a backdraft-predicting computer model. The NIST's team plans to incorporate the model into handheld devices that firefighters could use to take simple measurements through small openings in a room.

A lack of oxygen can reduce even the most furious flame to smoldering ash. But when fresh air rushes in, say after a firefighter opens a window or door to a room, the blaze may be suddenly and violently resurrected. This explosive phenomenon, called backdraft, can be lethal and has been challenging for firefighters to anticipate.

Now, researchers at the National Institute of Standards and Technology (NIST) have hatched a plan for informing firefighters of what dangers lie behind closed doors. The team obtained data from hundreds of backdrafts in the lab to use as a basis for a model that can predict backdrafts. The results of a new study, described at the 2022 Suppression, Detection and Signaling Research and Applications Conference, suggest that the model offers a viable solution to make predictions based on particular measurements. In the future, the team seeks to implement the technology into small-scale devices that firefighters could deploy in the field to avoid or adapt to dangerous conditions.

Currently, firefighters are looking for visual indicators of a potential backdraft, including soot-stained windows, smoke puffing through small openings and the absence of flames. If the cues are present, they may vent the room by creating holes in its ceiling to reduce their risk. If not, they may charge right in. Ultimately, first responders must rely on their eyes in a hazy environment to guess the correct action. And guessing wrong could come at a steep cost.

An AI model reveals how the body’s defense system recognizes skin cancer

Boosting the body’s own defense system has proven to be a particularly effective therapy for skin cancer.
Photo credit: National Cancer Institute

The artificial intelligence model could be utilized to enable more effective care for skin cancer patients and could lead to similar breakthroughs in the diagnosis and treatment of other cancers.

Researchers from the University of Helsinki, HUS Comprehensive Cancer Center, Aalto University and Stanford University have developed an artificial intelligence model that predicts which skin cancer patients will benefit from a treatment that activates the immune defense system. In practice, the AI model makes it possible to diagnose skin cancer with a blood test, determine the prognosis and target therapies increasingly accurately.

The skin cancer–related study was published in the esteemed Nature Communications journal.

The right medication for the right patient

Boosting the body’s own defense system has proven to be a particularly effective therapy for skin cancer. The problem with therapies that activate the immune system are the differences between patient groups: while some patients can be said to be cured, others gain no benefit from the treatment at all.

“Prior research has been unable to provide doctors with tools that would predict who will benefit from treatment that activates the defense system. The correct targeting of therapies is extremely important, since drug therapies are expensive and serious adverse effects fairly common,” says doctor and Doctoral Researcher Jani Huuhtanen from the University of Helsinki and Aalto University.

Team uses digital cameras, machine learning to predict neurological disease

From left, Richard Sowers, Rachneet Kaur and Manuel Hernandez led the development of a new approach for identifying people with multiple sclerosis or Parkinson’s disease. Their method involves videotaping the hips and lower extremities of individuals walking on a treadmill and allowing a machine-learning algorithm to differentiate gait abnormalities associated with each of these neurological conditions.
Photo credit: Fred Zwicky

In an effort to streamline the process of diagnosing patients with multiple sclerosis and Parkinson’s disease, researchers used digital cameras to capture changes in gait – a symptom of these diseases – and developed a machine-learning algorithm that can differentiate those with MS and PD from people without those neurological conditions.

Their findings are reported in the IEEE Journal of Biomedical and Health Informatics.

The goal of the research was to make the process of diagnosing these diseases more accessible, said Manuel Hernandez, a University of Illinois Urbana-Champaign professor of kinesiology and community health who led the work with graduate student Rachneet Kaur and industrial and enterprise systems engineering and mathematics professor Richard Sowers.

Currently, patients must wait – sometimes for years – to get an appointment with a neurologist to make a diagnosis, Hernandez said. And people in rural communities often must travel long distances to a facility where their condition can be assessed. To be able to gather gait information using nothing more than a digital camera and have that data assessed online could allow clinicians to do a quick screening that sends to a specialist only those deemed likely to have a neurological condition.

Claims AI can boost workplace diversity are ‘spurious and dangerous’, researchers argue

Co-author Dr Eleanor Drage testing the 'personality machine' built by Cambridge undergraduates.
  Credit: Eleanor Drage

Recent years have seen the emergence of AI tools marketed as an answer to lack of diversity in the workforce, from use of chatbots and CV scrapers to line up prospective candidates, through to analysis software for video interviews.

Those behind the technology claim it cancels out human biases against gender and ethnicity during recruitment, instead using algorithms that read vocabulary, speech patterns and even facial micro-expressions to assess huge pools of job applicants for the right personality type and “culture fit”.

However, in a new report published in Philosophy and Technology, researchers from Cambridge’s Centre for Gender Studies argue these claims make some uses of AI in hiring little better than an “automated pseudoscience” reminiscent of physiognomy or phrenology: the discredited beliefs that personality can be deduced from facial features or skull shape.

They say it is a dangerous example of “techno solutionism”: turning to technology to provide quick fixes for deep-rooted discrimination issues that require investment and changes to company culture.

Mathematical Formula Tackles Complex Moral Decision-Making in AI

Photo credit: Andy Kelly.

An interdisciplinary team of researchers has developed a blueprint for creating algorithms that more effectively incorporate ethical guidelines into artificial intelligence (AI) decision-making programs. The project was focused specifically on technologies in which humans interact with AI programs, such as virtual assistants or “carebots” used in healthcare settings.

“Technologies like carebots are supposed to help ensure the safety and comfort of hospital patients, older adults and other people who require health monitoring or physical assistance,” says Veljko Dubljević, corresponding author of a paper on the work and an associate professor in the Science, Technology & Society program at North Carolina State University. “In practical terms, this means these technologies will be placed in situations where they need to make ethical judgments.

“For example, let’s say that a carebot is in a setting where two people require medical assistance. One patient is unconscious but requires urgent care, while the second patient is in less urgent need but demands that the carebot treat him first. How does the carebot decide which patient is assisted first? Should the carebot even treat a patient who is unconscious and therefore unable to consent to receiving the treatment?

“Previous efforts to incorporate ethical decision-making into AI programs have been limited in scope and focused on utilitarian reasoning, which neglects the complexity of human moral decision-making,” Dubljević says. “Our work addresses this and, while I used carebots as an example, is applicable to a wide range of human-AI teaming technologies.”

As ransomware attacks increase, new algorithm may help prevent power blackouts

Saurabh Bagchi, a Purdue professor of electrical and computer engineering, develops ways to improve the cybersecurity of power grids and other critical infrastructure.
Credit: Purdue University photo/Vincent Walter

Millions of people could suddenly lose electricity if a ransomware attack just slightly tweaked energy flow onto the U.S. power grid.

No single power utility company has enough resources to protect the entire grid, but maybe all 3,000 of the grid’s utilities could fill in the most crucial security gaps if there were a map showing where to prioritize their security investments.

Purdue University researchers have developed an algorithm to create that map. Using this tool, regulatory authorities or cyber insurance companies could establish a framework that guides the security investments of power utility companies to parts of the grid at greatest risk of causing a blackout if hacked.

Power grids are a type of critical infrastructure, which is any network – whether physical like water systems or virtual like health care record keeping – considered essential to a country’s function and safety. The biggest ransomware attacks in history have happened in the past year, affecting most sectors of critical infrastructure in the U.S. such as grain distribution systems in the food and agriculture sector and the Colonial Pipeline, which carries fuel throughout the East Coast.

Repurposing existing drugs to fight new COVID-19 variants

Photo Credit: Myriam Zilles

MSU researchers are using big data and AI to identify current drugs that could be applied to treat new COVID-19 variants

Finding new ways to treat the novel coronavirus and its ever-changing variants has been a challenge for researchers, especially when the traditional drug development and discovery process can take years. A Michigan State University researcher and his team are taking a hi-tech approach to determine whether drugs already on the market can pull double duty in treating new COVID variants.

“The COVID-19 virus is a challenge because it continues to evolve,” said Bin Chen, an associate professor in the College of Human Medicine. “By using artificial intelligence and really large data sets, we can repurpose old drugs for new uses.”

Chen built an international team of researchers with expertise on topics ranging from biology to computer science to tackle this challenge. First, Chen and his team turned to publicly available databases to mine for the unique coronavirus gene expression signatures from 1,700 host transcriptomic profiles that came from patient tissues, cell cultures and mouse models. These signatures revealed the biology shared by COVID-19 and its variants.

Scientists chart how exercise affects the body

MIT and Harvard Medical School researchers mapped out many of the cells, genes, and cellular pathways that are modified by exercise or high-fat diet.
Photo Credit: Gabin Vallet

Exercise is well-known to help people lose weight and avoid gaining it. However, identifying the cellular mechanisms that underlie this process has proven difficult because so many cells and tissues are involved.

In a new study in mice that expands researchers’ understanding of how exercise and diet affect the body, MIT and Harvard Medical School researchers have mapped out many of the cells, genes, and cellular pathways that are modified by exercise or high-fat diet. The findings could offer potential targets for drugs that could help to enhance or mimic the benefits of exercise, the researchers say.

“It is extremely important to understand the molecular mechanisms that are drivers of the beneficial effects of exercise and the detrimental effects of a high-fat diet, so that we can understand how we can intervene, and develop drugs that mimic the impact of exercise across multiple tissues,” says Manolis Kellis, a professor of computer science in MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and a member of the Broad Institute of MIT and Harvard.

The researchers studied mice with high-fat or normal diets, who were either sedentary or given the opportunity to exercise whenever they wanted. Using single-cell RNA sequencing, the researchers cataloged the responses of 53 types of cells found in skeletal muscle and two types of fatty tissue.

Mouse-human comparison shows unimagined functions of the Thalamus

With mathematical models, Bochum and US researchers have simulated processes in the brain of mice and humans.
Credit: RUB, Marquard

Researchers have reproduced the brain functions of the mouse and human in the computer. Artificial intelligence could learn from this.

For a long time, the thalamus was considered a brain region that is primarily responsible for processing sensory stimuli. Current studies have increased the evidence that it is a central switch in cognitive processes. Researchers of neuroscience around Prof. Dr. Burkhard Pleger in Collaborative Research Center 874 of the Ruhr University Bochum and a team from the Massachusetts Institute of Technology (MIT, USA) observed learning processes in the brains of mice and humans and reproduced them in mathematical models. They were able to show that the region of the mediodoral nucleus in the thalamus has a decisive share in cognitive flexibility. They report in the journal PLOS Computational Biology.

AI boosts usability of paper-making waste products

Photo and graphic with birch tree by J. Löfgren

In a new and exciting collaboration with the Department of Bioproducts and Biosystems, researchers in the CEST group have published a study demonstrating how artificial intelligence (AI) can boost the production of renewable biomaterials. Their publication focuses on the extraction of lignin, an organic polymer that together with cellulose makes up the cell walls of plants. As a side-product of papermaking, lignin is produced in large quantities around the world but seldom used as anything other than cheap fuel. Developing valuable materials and chemicals from lignin would consequently be a big step towards a sustainable society.

A key challenge for the valorization of lignin is to find the right experimental extraction conditions. These include things like the temperature in the hot-water reactor where the wood is processed, the reaction time and the ratio of wood to water. These conditions not only affect the amount of lignin that can be extracted, but also the physical and chemical properties of the extracted lignin itself. Therefore, knowing how to choose the right experimental conditions is important since the more lignin can be extracted the better, and different lignin-based products may require lignin with different properties.

Machine learning may enable bioengineering of the most abundant enzyme

Photo Credit: Melissa Askew

A Newcastle University study has for the first time shown that machine learning can predict the biological properties of the most abundant enzyme on Earth - Rubisco.

Rubisco (Ribulose-1,5-bisphosphate carboxylase/oxygenase) is responsible for providing carbon for almost all life on Earth. Rubisco functions by converting atmospheric CO2 from the Earth’s atmosphere to organic carbon matter, which is essential to sustain most life on Earth.

For some time now, natural variation has been observed among Rubisco proteins of land plants and modelling studies have shown that transplanting Rubisco proteins with certain functional properties can increase the amount of atmospheric CO2 crop plants can uptake and store.

Study lead author, Wasim Iqbal, a PhD researcher at Newcastle University’s School of Natural and Environmental Sciences, part of Dr Maxim Kapralov’s group, developed a machine learning tool which can predict the performance properties of numerous land plant Rubisco proteins with surprisingly good accuracy. The hope is that this tool will enable the hunt for a ‘supercharged’ Rubisco protein that can be bioengineered into major crops such as wheat.

Neurodegenerative disease can progress in newly identified patterns

Neurodegenerative diseases — like amyotrophic lateral sclerosis (ALS, or Lou Gehrig's disease), Alzheimer’s, and Parkinson’s — are complicated, chronic ailments that can present with a variety of symptoms, worsen at different rates, and have many underlying genetic and environmental causes, some of which are unknown. ALS, in particular, affects voluntary muscle movement and is always fatal, but while most people survive for only a few years after diagnosis, others live with the disease for decades. Manifestations of ALS can also vary significantly; often slower disease development correlates with onset in the limbs and affecting fine motor skills, while the more serious, bulbar ALS impacts swallowing, speaking, breathing, and mobility. Therefore, understanding the progression of diseases like ALS is critical to enrollment in clinical trials, analysis of potential interventions, and discovery of root causes.

However, assessing disease evolution is far from straightforward. Current clinical studies typically assume that health declines on a downward linear trajectory on a symptom rating scale, and use these linear models to evaluate whether drugs are slowing disease progression. However, data indicate that ALS often follows nonlinear trajectories, with periods where symptoms are stable alternating with periods when they are rapidly changing. Since data can be sparse, and health assessments often rely on subjective rating metrics measured at uneven time intervals, comparisons across patient populations are difficult. These heterogenous data and progression, in turn, complicate analyses of invention effectiveness and potentially mask disease origin.

Conventional Computers Can Learn to Solve Tricky Quantum Problems

Hsin-Yuan (Robert) Huang
Credit: Caltech

There has been a lot of buzz about quantum computers and for good reason. The futuristic computers are designed to mimic what happens in nature at microscopic scales, which means they have the power to better understand the quantum realm and speed up the discovery of new materials, including pharmaceuticals, environmentally friendly chemicals, and more. However, experts say viable quantum computers are still a decade away or more. What are researchers to do in the meantime?

A new Caltech-led study in the journal Science describes how machine learning tools, run on classical computers, can be used to make predictions about quantum systems and thus help researchers solve some of the trickiest physics and chemistry problems. While this notion has been proposed before, the new report is the first to mathematically prove that the method works in problems that no traditional algorithms could solve.

"Quantum computers are ideal for many types of physics and materials science problems," says lead author Hsin-Yuan (Robert) Huang, a graduate student working with John Preskill, the Richard P. Feynman Professor of Theoretical Physics and the Allen V. C. Davis and Lenabelle Davis Leadership Chair of the Institute for Quantum Science and Technology (IQIM). "But we aren't quite there yet and have been surprised to learn that classical machine learning methods can be used in the meantime. Ultimately, this paper is about showing what humans can learn about the physical world."

Wildfire smoke is unraveling decades of air quality gains

Over the last decade, PM2.5 from wildfire smoke has increased in much of the U.S., particularly in Western states, but some areas in the South and East have seen modest declines. This map shows the decadal change in smoke PM2.5, meaning the difference in daily average smoke PM2.5 during 2006−2010 compared to 2016−2020.
Image credit: Childs et al. 2022, Environmental Science & Technology

Stanford researchers have developed an AI model for predicting dangerous particle pollution to help track the American West’s rapidly worsening wildfire smoke. The detailed results show millions of Americans are routinely exposed to pollution at levels rarely seen just a decade ago.

Wildfire smoke now exposes millions of Americans each year to dangerous levels of fine particulate matter, lofting enough soot across parts of the West in recent years to erase much of the air quality gains made over the last two decades.

Those are among the findings of a new Stanford University study published Sept. 22 in Environmental Science & Technology that focuses on a type of particle pollution known as PM2.5, which can lodge deep in our lungs and even get into our bloodstream.

Using statistical modeling and artificial intelligence techniques, the researchers estimated concentrations of PM2.5 specifically from wildfire smoke in sharp enough detail to reveal variations within individual counties and individual smoke events from coast to coast from 2006 to 2020.

“We found that people are being exposed to more days with wildfire smoke and more extreme days with high levels of fine particulate matter from smoke,” said lead study author Marissa Childs, who worked on the research as a PhD student in Stanford’s Emmett Interdisciplinary Program in Environment and Resources (E-IPER). Unlike other major pollutant sources, wildfire smoke is considered an “exceptional event” under the Clean Air Act, she explained, “which means an increasing portion of the particulate matter that people are exposed to is unregulated.”

Shutting down backup genes leads to cancer remission in mice

Abhinav Achreja, PhD, Research Fellow at the University of Michigan Biomedical Engineering and Deepak Nagrath, Ph.D. Associate Professor of Biomedical Engineering works on ovarian cancer cell research in the bio-engineering lab at the North Campus Research Center (NCRC).
Image credit: Marcin Szczepanski, Michigan Engineering

The way that tumor cells enable their uncontrolled growth is also a weakness that can be harnessed to treat cancer, researchers at the University of Michigan and Indiana University have shown.

Their machine-learning algorithm can identify backup genes that only tumor cells are using so that drugs can target cancer precisely.

“Most cancer drugs affect normal tissues and cells. However, our strategy allows specific targeting of cancer cells.”

Deepak Nagrath

The team demonstrated this new precision medicine approach for treating ovarian cancer in mice. Moreover, the cellular behavior that exposes these vulnerabilities is common across most forms of cancer, meaning the algorithms could provide better treatment plans for a host of malignancies.

“This could revolutionize the precision medicine field because the drug targeting will only affect and kill cancer cells and spare the normal cells,” said Deepak Nagrath, a U-M associate professor of biomedical engineering and senior author of the study in Nature Metabolism. “Most cancer drugs affect normal tissues and cells. However, our strategy allows specific targeting of cancer cells.”

Even the smartest AI models don’t match human visual processing

The study employed novel visual stimuli called “Frankensteins
Source/Credit: York University

Deep convolutional neural networks (DCNNs) don’t see objects the way humans do – using configural shape perception – and that could be dangerous in real-world AI applications, says Professor James Elder, co-author of a York University study.

Published in the Cell Press journal iScience, Deep learning models fail to capture the configural nature of human shape perception is a collaborative study by Elder, who holds the York Research Chair in Human and Computer Vision and is Co-Director of York’s Centre for AI & Society, and Assistant Psychology Professor Nicholas Baker at Loyola College in Chicago, a former VISTA postdoctoral fellow at York.

The study employed novel visual stimuli called “Frankensteins” to explore how the human brain and DCNNs process holistic, configural object properties.

“Frankensteins are simply objects that have been taken apart and put back together the wrong way around,” says Elder. “As a result, they have all the right local features, but in the wrong places.”

The investigators found that while the human visual system is confused by Frankensteins, DCNNs are not – revealing an insensitivity to configural object properties.

New method for comparing neural networks exposes how artificial intelligence works

Researchers at Los Alamos are looking at new ways to compare neural networks. This image was created with an artificial intelligence software called Stable Diffusion, using the prompt “Peeking into the black box of neural networks.”
Source: Los Alamos National Laboratory

A team at Los Alamos National Laboratory has developed a novel approach for comparing neural networks that looks within the “black box” of artificial intelligence to help researchers understand neural network behavior. Neural networks recognize patterns in datasets; they are used everywhere in society, in applications such as virtual assistants, facial recognition systems and self-driving cars.

“The artificial intelligence research community doesn’t necessarily have a complete understanding of what neural networks are doing; they give us good results, but we don’t know how or why,” said Haydn Jones, a researcher in the Advanced Research in Cyber Systems group at Los Alamos. “Our new method does a better job of comparing neural networks, which is a crucial step toward better understanding the mathematics behind AI.”

Jones is the lead author of the paper “If You’ve Trained One You’ve Trained Them All: Inter-Architecture Similarity Increases With Robustness,” which was presented recently at the Conference on Uncertainty in Artificial Intelligence. In addition to studying network similarity, the paper is a crucial step toward characterizing the behavior of robust neural networks.

New AI system predicts how to prevent wildfires

Satellite image of Borneo in 2006 covered by smoke from fires (marked by red dots).
Image Credit: Jeff Schmaltz, MODIS Rapid Response Team / NASA

A machine learning model can evaluate the effectiveness of different management strategies

Wildfires are a growing threat in a world shaped by climate change. Now, researchers at Aalto University have developed a neural network model that can accurately predict the occurrence of fires in peatlands. They used the new model to assess the effect of different strategies for managing fire risk and identified a suite of interventions that would reduce fire incidence by 50-76%.

The study focused on the Central Kalimantan province of Borneo in Indonesia, which has the highest density of peatland fires in Southeast Asia. Drainage to support agriculture or residential expansion has made peatlands increasingly vulnerable to recurring fires. In addition to threatening lives and livelihoods, peatland fires release significant amounts of carbon dioxide. However, prevention strategies have faced difficulties because of the lack of clear, quantified links between proposed interventions and fire risk.

The new model uses measurements taken before each fire season in 2002-2019 to predict the distribution of peatland fires. While the findings can be broadly applied to peatlands elsewhere, a new analysis would have to be done for other contexts. ‘Our methodology could be used for other contexts, but this specific model would have to be re-trained on the new data,’ says Alexander Horton, the postdoctoral researcher who carried out study.