Aurelia Butler

Computers work in digits — 0s and 1s to be exact. Their calculations are digital; their processes are digital; even their memories are digital. All of which requires extraordinary power resources. As we look to the next evolution of computing and developing neuromorphic or “brain like” computing, those power requirements are unfeasible.
To advance neuromorphic computing, some researchers are looking at analog improvements. In other words, not just advancing software, but advancing hardware too. Research from the University of California San Diego and UC Riverside shows a promising new way to store and transmit information using disordered superconducting loops.
The team’s research, which appears in the Proceedings of the National Academy of Sciences, offers the ability of superconducting loops to demonstrate associative memory, which, in humans, allows the brain to remember the relationship between two unrelated items.
“I hope what we’re designing, simulating and building will be able to do that kind of associative processing really fast,” stated UC San Diego Professor of Physics Robert C. Dynes, who is one of the paper’s co-authors.
Creating lasting memories
Picture it: you’re at a party and run into someone you haven’t seen in a while. You know their name but can’t quite recall it. Your brain starts to root around for the information: where did I meet this person? How were we introduced? If you’re lucky, your brain finds the pathway to retrieve what was missing. Sometimes, of course, you’re unlucky.
Dynes believes that short-term memory moves into long-term memory with repetition. In the case of a name, the more you see the person and use the name, the more deeply it is written into memory. This is why we still remember a song from when we were ten years old but can’t remember what we had for lunch yesterday.

“Our brains have this remarkable gift of associative memory, which we don’t really understand,” stated Dynes, who is also president emeritus of the University of California and former UC San Diego chancellor. “It can work through the probability of answers because it’s so highly interconnected. This computer brain we built and modeled is also highly interactive. If you input a signal, the whole computer brain knows you did it.”
Staying in the loop
How do disordered superconducting loops work? You need a superconducting material — in this case, the team used yttrium barium copper oxide (YBCO). Known as a high-temperature superconductor, YBCO becomes superconducting around 90 Kelvin (-297 F), which in the world of physics, is not that cold. This made it relatively easy to modify. The YBCO thin films (about 10 microns wide) were manipulated with a combination of magnetic fields and currents to create a single flux quantum on the loop. When the current was removed, the flux quantum stayed in the loop. Think of this as a piece of information or memory.
This is one loop, but associative memory and processing require at least two pieces of information. For this, Dynes used disordered loops, meaning the loops are different sizes and follow different patterns — essentially random.
A Josephson juncture, or “weak link,” as it is sometimes known, in each loop acted as a gate through which the flux quanta could pass. This is how information is transferred and the associations are built.
Although traditional computing architecture has continuous high-energy requirements, not just for processing but also for memory storage, these superconducting loops show significant power savings — on the scale of a million times less. This is because the loops only require power when performing logic tasks. Memories are stored in the physical superconducting material and can remain there permanently, as long as the loop remains superconducting.

The number of memory locations available increases exponentially with more loops: one loop has three locations, but three loops have 27. For this research, the team built four loops with 81 locations. Next, Dynes would like to expand the number of loops and the number memory locations.
“We know these loops can store memories. We know the associative memory works. We just don’t know how stable it is with a higher number of loops,” he said.
This work is not only noteworthy to physicists and computer engineers; it may also be important to neuroscientists. Dynes talked to another University of California president emeritus, Richard Atkinson, a world-renowned cognitive scientist who helped create a seminal model of human memory called the Atkinson-Shiffrin model.
Atkinson, who is also former UC San Diego chancellor and professor emeritus in the School of Social Sciences, was excited about the possibilities he saw: “Bob and I have had some great discussions trying to determine if his physics-based neural network could be used to model the Atkinson-Shiffrin theory of memory. His system is quite different from other proposed physics-based neural networks, and is rich enough that it could be used to explain the workings of the brain’s memory system in terms of the underlying physical process. It’s a very exciting prospect.”
Full list of authors: Uday S. Goteti and Robert C. Dynes (both UC San Diego); Shane A. Cybart (UC Riverside).
This work was primarily supported as part of the Quantum Materials for Energy Efficient Neuromorphic Computing (Q-MEEN-C) (Department of Energy DE-SC0019273). Other support was provided by the Department of Energy National Nuclear Security Agency (DE-NA0004106) and the Air Force Office of Scientific Research (FA9550-20-1-0144). More

Through steady advances in the development of quantum computers and their ever-improving performance, it will be possible in the future to crack our current encryption processes. To address this challenge, researchers at the Technical University of Munich (TUM) are participating in an international research consortium to develop encryption methods that will apply physical laws to prevent the interception of messages. To safeguard communications over long distances, the QUICK³ space mission will deploy satellites.
How can it be ensured that data transmitted through the internet can be read only by the intended recipient? At present our data are encrypted with mathematical methods that rely on the idea that the factorization of large numbers is a difficult task. With the increasing power of quantum computers, however, these mathematical codes will probably no longer be secure in the future.
Encryption by means of physical laws
Tobias Vogl, a professor of Quantum Communication Systems Engineering, is working on an encryption process that relies on principles of physics. “Security will be based on the information being encoded into individual light particles and then transmitted. The laws of physics do not permit this information to be extracted or copied. When the information is intercepted, the light particles change their characteristics. Because we can measure these state changes, any attempt to intercept the transmitted data will be recognized immediately, regardless of future advances in technology,” says Tobias Vogl.
The big challenge in so-called quantum cryptography lies in the transmission of data over long distances. In classical communications, information is encoded in many light particles and transmitted through optical fibers. However, the information in a single particle cannot be copied. As a result, the light signal cannot be repeatedly amplified, as with current optical fiber transmissions. This limits the transmission distance for the information to a few hundred kilometers.
To send information to other cities or continents, the structure of the atmosphere will be used. At altitudes higher than around 10 kilometers, the atmosphere is so thin that light is neither scattered nor absorbed. This will make it possible to use satellites in order to extend quantum communications over longer distances.
Satellites for quantum communications
As part of the QUICK³ mission, Tobias Vogl and his team are developing an entire system, including all of the components needed to build a satellite for quantum communications. In a first step, the team tested each of the satellite components. The next step will be to try out the entire system in space. The researchers will investigate whether the technology can withstand outer space conditions and how the individual system components interact. The satellite launch is scheduled for 2025. To create an overarching network for quantum communications, however, hundreds or perhaps thousands of satellites will be needed.
Hybrid network for encryption
The concept does not necessarily require all information to be transmitted using this method, which is highly complex and costly. It is conceivable that a hybrid network could be implemented in which data can be encrypted either physically or mathematically. Antonia Wachter-Zeh, a professor of Coding and Cryptography, is working to develop algorithms sufficiently complex that not even quantum computers can solve them. In the future it will still be enough to encrypt most information using mathematical algorithms. Quantum cryptography will be an option only for documents requiring special protection, for example in communications between banks. More

Scientists from Ames National Laboratory have identified the first unconventional superconductor with a chemical composition also found in nature. Miassite is one of only four minerals found in nature that act as a superconductor when grown in the lab. The team’s investigation of miassite revealed that it is an unconventional superconductor with properties similar to high-temperature superconductors. Their findings further scientists’ understanding of this type of superconductivity, which could lead to more sustainable and economical superconductor-based technology in the future.
Superconductivity is when a material can conduct electricity without energy loss. Superconductors have applications including medical MRI machines, power cables, and quantum computers. Conventional superconductors are well understood but have low critical temperatures. The critical temperature is the highest temperature at which a material acts as a superconductor.
In the 1980s, scientists discovered unconventional superconductors, many of which have much higher critical temperatures. According to Ruslan Prozorov, a scientist at Ames Lab, all these materials are grown in the lab. This fact has led to the general belief that unconventional superconductivity is not a natural phenomenon.
Prozorov explained that it is difficult to find superconductors in nature because most superconducting elements and compounds are metals and tend to react with other elements, like oxygen. He said that miassite (Rh17S15) is an interesting mineral for several reasons, one of which is its complex chemical formula. “Intuitively, you think that this is something which is produced deliberately during a focused search, and it cannot possibly exist in nature,” said Prozorov, “But it turns out it does.”
Paul Canfield, Distinguished Professor of Physics and Astronomy at Iowa State University and a scientist at Ames Lab, has expertise in design, discovery, growth, and characterization of novel crystalline materials. He synthesized high quality miassite crystals for this project. “Although miassite is a mineral that was discovered near the Miass River in Chelyabinsk Oblast, Russia,” said Canfield, “it is a rare one that generally does not grow as well-formed crystals.”
Growing the miassite crystals was part of a larger effort to discover compounds that combine very high melting elements (like Rh) and volatile elements (like S). “Contrary to the nature of the pure elements, we have been mastering the use of mixtures of these elements that allow for low temperature growth of crystals with minimal vapor pressure,” said Canfield. “It’s like finding a hidden fishing hole that is full of big fat fish. In the Rh-S system we discovered three new superconductors. And, through Ruslan’s detailed measurements, we discovered that the miassite is an unconventional superconductor.”
Prozorov’s group specializes in advanced techniques to study superconductors at low temperatures. He said the material needed to be as cold as 50 millikelvins, which is about -460 degrees Fahrenheit.

Prozorov’s team used three different tests to determine the nature of miassite’s superconductivity. The main test is called the “London penetration depth.” It determines how far a weak magnetic field can penetrate the superconductor bulk from the surface. In a conventional superconductor, this length is basically constant at low temperature. However, in unconventional superconductors, it varies linearly with the temperature. This test showed that miassite behaves as an unconventional superconductor.
Another test the team performed was introducing defects into the material. Prozorov said that this test is a signature technique his team has employed over the past decade. It involves bombarding the material with high-energy electrons. This process knocks-out ions from their positions, thus creating defects in the crystal structure. This disorder can cause changes in the material’s critical temperature.
Conventional superconductors are not sensitive to non-magnetic disorder, so this test would show no or very little change in the critical temperature. Unconventional superconductors have a high sensitivity to disorder, and introducing defects changes or suppresses the critical temperature. It also affects the critical magnetic field of the material. In miassite, the team found that both the critical temperature and the critical magnetic field behaved as predicted in unconventional superconductors.
Investigating unconventional superconductors improves scientists understanding of how they work. Prozorov explained that this is important because, “Uncovering the mechanisms behind unconventional superconductivity is key to economically sound applications of superconductors.” More

A new study published today (Tuesday 12 March) by the University of Sussex shows how researchers are using AI technology and social media to help identify global threats to wildlife.
Researchers at Sussex have used AI to access online records from Facebook, X/Twitter, Google, and Bing, to map the global extent of threats to bats from hunting and trade.
The new study demonstrates how social media and online content generated by news outlets and the public, can help to increase our understanding of threats to wildlife across the world — and refocus conservation efforts.
The Sussex team identified 22 countries involved in bat exploitation, covering both hunting and trade, that had not previously been identified by traditional academic research, including Bahrain, Spain, Sri Lanka, New Zealand and Singapore, which had the highest number of new records.
The team developed an automated system which allowed them to conduct large scale searches across multiple platforms. Using AI, they filtered tens of thousands of results to find relevant data. Any observations or anecdotes of bat exploitation were used to develop a global database of ‘bat exploitation records’.
To better understand threats to bats, the team compared online records with academic records, knowing that data and information shared online is influenced by factors including global events and where people have access to the internet.
Lead author, Bronwen Hunter at the University of Sussex says:
“Using data sources like this provides a low-cost way to help us understand threats to wildlife globally. AI allowed us to access the data at scale and complete a global analysis, which isn’t something we would have been able to achieve using traditional field studies.

“Another benefit of using online data combined with automated data filtering is that more information can be obtained in real-time, ensuring that we can keep up to date with current threats.”
Bats make up about a fifth of all mammal species globally, and have a vital role in ecosystems. They are pollinators, disperse seeds and help with pest control.
Over half of bat species are considered as either ‘Threatened with Extinction’ or ‘Data Deficient’ by the International Union for Nature Conservation (IUCN). Much less is known about the impact of hunting and trade of bats compared with other mammals. However, their very low reproductive rate and longevity — usually 10-30 years — makes them likely to be vulnerable on a scale more commonly associated with much larger mammals such as chimpanzees, bears or lions.
Being able to expand knowledge of bat exploitation using crowd-sourced digital records can help identify bat populations most in need of conservation action, or feed into global assessments, such as the IUCN Red List.
Prof. Fiona Mathews at the University of Sussex, who leads the research group says:
“The hunting and sale of bats for meat was highlighted during the Covid pandemic. But there is also a worrying trade of bats as curios or medicines. It is vital that we understand where bat exploitation is happening, and this has been very difficult historically because it often happens in remote places, and elicit trade can be hidden. This research shows that posts on the internet and social media can provide vital evidence, that can now be followed up on the ground.”
This research highlights the value of contributions from social media and online platforms and argues that they could be used for future conservation decision making. Using online data combined with current research studies provides a more complete picture of the global extent of bat exploitation.

Kit Stoner, CEO at The Bat Conservation Trust says:
“Unsustainable wildlife trade can pose a threat to bat species being hunted or harvested. Often, species are sold much further afield from where they are found. This trade can undermine bat conservation directly and pose a wider threat in terms of increasing the risk of zoonosis. We welcome the results of this research in providing a possible new low-cost way of detecting trade in bats which could offer a way of monitoring how this wildlife trade operates and examining ways of disrupting it.” More

It might look like a roll of chicken wire, but this tiny cylinder of carbon atoms — too small to see with the naked eye — could one day be used for making electronic devices ranging from night vision goggles and motion detectors to more efficient solar cells, thanks to techniques developed by researchers at Duke University.
First discovered in the early 1990s, carbon nanotubes are made from single sheets of carbon atoms rolled up like a straw.
Carbon isn’t exactly a newfangled material. All life on Earth is based on carbon. It’s the same stuff found in diamonds, charcoal, and pencil lead.
What makes carbon nanotubes special are their remarkable properties. These tiny cylinders are stronger than steel, and yet so thin that 50,000 of them would equal the thickness of a human hair.
They’re also amazingly good at conducting electricity and heat, which is why, in the push for faster, smaller, more efficient electronics, carbon nanotubes have long been touted as potential replacements for silicon.
But producing nanotubes with specific properties is a challenge.
Depending on how they’re rolled up, some nanotubes are considered metallic — meaning electrons can flow through them at any energy. The problem is they can’t be switched off. This limits their use in digital electronics, which use electrical signals that are either on or off to store binary states; just like silicon semiconductor transistors switch between 0 and 1 bits to carry out computations.

Duke chemistry professor Michael Therien and his team say they’ve found a way around this.
The approach takes a metallic nanotube, which always lets current through, and transforms it into a semiconducting form that can be switched on and off.
The secret lies in special polymers — substances whose molecules are hooked together in long chains — that wind around the nanotube in an orderly spiral, “like wrapping a ribbon around a pencil,” said first author Francesco Mastrocinque, who earned his chemistry Ph.D. in Therien’s lab at Duke.
The effect is reversible, they found. Wrapping the nanotube in a polymer changes its electronic properties from a conductor to a semiconductor. But if the nanotube is unwrapped, it goes back to its original metallic state.
The researchers also showed that by changing the type of polymer that encircles a nanotube, they could engineer new types of semiconducting nanotubes. They can conduct electricity, but only when the right amount of external energy is applied.
“This method provides a subtle new tool,” Therien said. “It allows you to make a semiconductor by design.”
Practical applications of the method are likely far off. “We’re a long way from making devices,” Therien said.

Mastrocinque and his co-authors say the work is important because it’s a way to design semiconductors that can conduct electricity when struck by light of certain low-energy wavelengths that are common but invisible to human eyes.
In the future for instance, the Duke team’s work might help others engineer nanotubes that detect heat released as infrared radiation, to reveal people or vehicles hidden in the shadows. When infrared light — such as that emitted by warm-blooded animals — strikes one of these nanotube-polymer hybrids, it would generate an electric signal.
Or take solar cells: this technique could be used to make nanotube semiconductors that convert a broader range of wavelengths into electricity, to harness more of the Sun’s energy.
Because of the spiral wrapper on the nanotube surface, these structures could also be ideal materials for new forms of computing and data storage that use the spins of electrons, in addition to their charge, to process and carry information.
The researchers describe their results March 11 in the journal Proceedings of the National Academy of Sciences.
This research was supported by the Air Force Office of Scientific Research (FA9550-18-1-0222), the National Institutes of Health (1R01HL146849), the United States National Science Foundation (CHE-2140249, DGE-2040435) and the John Simon Guggenheim Memorial Foundation. More

Numerous measurement stations around the world provide us with data about air quality, allowing us to enhance it. Although we are increasingly collecting data from marine areas, access to such data is considerably more challenging. Signals are poorly transmitted through water, differences in both pressure and currents hinder measurement devices and there is an absence of pre-constructed computing infrastructure.
Could intelligent technologies help us improve marine data collection? Professor of Computer Science Petteri Nurmi and his research group at the University of Helsinki have joined forces with researchers at the University of Tartu, University of Madeira, and MARE-Madeira, ARDITI, a non-profit marine research institute, to develop solutions combining sensor technologies and embedded Artificial Intelligence.
The researchers strive to make more efficient and expansive the data collection methods now used in, for example, environmental research.
“The higher the quantity and quality of data about the oceans obtained, the better we can use it to understand and protect the oceans. Our methods help expand the total amount of data gathered from marine areas and reduce the effort required to collect and analyse them,” Nurmi says.
AI identifies animal species
In a recently published study, Nurmi and his colleagues used data collected from whale-watching excursions in Madeira, Portugal. The vessels used for these excursions usually carry persons who can record observations of the species seen or video the environment during the tour.
In the study, an AI model assisted individuals in environmental real-time observation. AI was also used to identify whether the video footage showed certain animals, such as dolphins or whales. In addition, the researchers compared the AI assistance received by experienced and less experienced observers, and explored how AI-assisted animal observations served as data for training the model.

“We analysed how AI assistance affected the quality of data and human observations. AI improved the accuracy of animal observations by amateurs, but had no effect on expert observations. On the other hand, when the data collected were used for training purposes, the best results were achieved by combining AI classifications with expert observations. Thus, interactions between humans and AI can influence each other and they need to be better understood,” Nurmi explains. The method could ideally be used for faster identification of animals moving in marine areas. The results and methods can also be expanded to observe other organisms.
Identifying marine plastics
Another recent study focused on identifying and classifying underwater plastic debris. These differ from microplastics, or tiny plastic particles, in being visible to the naked eye. One method currently used to obtain information on marine plastic pollution is to have divers or devices collect samples for laboratory analysis, but this usually takes a great deal of time. Surface-layer plastics can be observed with aerial photography as well.
The researchers developed an AI model functioning with sensors and based on the analysis of light spectrum data. The model could be connected to diver equipment or a diving robot to determine the type of plastic waste underwater. The researchers discovered that the model was able to distinguish types of plastic with 85% accuracy.
“We are capable of identifying four out of five objects directly, which means we need to send fewer samples to a lab for identification. This provides us with more data and thus a more comprehensive overview of marine plastic pollution.”
Nurmi emphasises that his and his colleagues’ aim is to create new ways of collecting data. Experts in other fields can then come up with the best uses for the new methods. For instance, once the type of plastic is known, it is easier to find out its source and consider ways of preventing pollution. The data also help researchers understand the effects on ecosystems of different types of plastic, as they contain different chemicals and break down in different ways.

Better data for better protection
Nurmi believes that the situation with marine data may be similar to the one in air quality research years ago.
“Originally, air quality research relied on a few large measuring towers, but now you can find them even on bus stops. In marine sciences this is not yet the case, even if the methods for collecting marine data and the number of actors involved are increasing. Our research ensures that we can obtain even greater amount of even more accurate marine measurement data as the capabilities to collect data improve.”
In the future, the use of underwater drones, smart buoys, vessels and coastal base stations is likely to increase in marine data collection. The expanding network of such tools will offer further ways to gather underwater data.
“Despite their vital importance for humanity, marine protection and regulation unfortunately tend to be drowned out by other issues. The higher the quantity and quality of marine data collected, the better we are able to develop solutions and regulations improving the status of the marine environment.” More

Research in quantum materials is paving the way for groundbreaking discoveries and is poised to drive technological advancements that will redefine the landscapes of industries like mining, energy, transportation, and medtech.
A technique called time- and angle-resolved photoemission spectroscopy (TR-ARPES) has emerged as a powerful tool, allowing researchers to explore the equilibrium and dynamical properties of quantum materials via light-matter interaction.
Published in the physics review journal, Reviews of Modern Physics, a recent review paper by Professor Fabio Boschini from the Institut national de la recherche scientifique (INRS), along with colleagues Marta Zonno from Canadian Light Source (CLS) and Andrea Damascelli from UBC’s Stewart Blusson Quantum Matter Institute (Blusson QMI), illustrates that TR-ARPES has rapidly matured into a powerful technique over the last two decades.
“TR-ARPES is an effective technique not only for fundamental studies, but also for characterizing out-of-equilibrium properties of quantum materials for future applications,” says Professor Boschini who specializes in ultrafast spectroscopies of condensed matter, at the Énergie Matériaux Télécommunications Research Centre.
A revolutionary tool for quantum materials research
The new paper provides a comprehensive review of research using TR-ARPES and its evolving significance in exploring light-induced electron dynamics and phase transitions in a wide range of quantum materials.
“The scientific community is currently investigating new ‘tuning knobs’ to control the electronic, transport, and magnetic properties of quantum materials on demand. One of these ‘tuning knobs’ is the light-matter interaction, which promises to provide fine control the properties of quantum materials on ultrafast timescales,” says Professor Boschini, who is also a QMI affiliate investigator. “TR-ARPES is the ideal technique for this purpose, since it provides direct insight into how light excitation modifies electronic states with time, energy, and momentum resolution.”
“TR-ARPES has ushered in a new era of quantum materials research, allowing us to ‘knock on the system’ and observe how it responds, and pushing the materials out of equilibrium to uncover their hidden properties,” adds Blusson QMI Scientific Director Andrea Damascelli.

Collaboration at the heart of TR ARPES’ success
TR-ARPES combines condensed matter spectroscopy (ARPES) with ultrafast lasers (photonics), bringing together research groups from both fields. The technique owes much of its success to significant advancements in developing new laser sources capable of producing light with precise characteristics.
Boschini is working closely on the subject with Professor François Légaré, a full professor at INRS and an expert in ultrafast laser science and technology. Together, Boschini’s and Légaré’s groups built and are operating a state-of-the-art TR-ARPES endstation with unique intense long-wavelength excitation capabilities at the Advanced Laser Light Source (ALLS) laboratory.
“Thanks to the support from the Canada Foundation for Innovation (CFI), the governments of Québec (MEIE) and Canada, and LaserNetUS, as well as the recent CFI Major Science Initiatives program, we are now in the privileged position to open the TR-ARPES endstation at ALLS to national and international users,” states Professor Légaré, Director of the Énergie Matériaux Télécommunications Research Centre at INRS and Scientific Head of ALLS.
According to Professor Boschini, TR-ARPES is now a mature technique with a proven impact on various branches of physics and chemistry. “Further experimental and theoretical developments, similar to what we are doing at ALLS, suggest that even more exciting times lie ahead,” he concludes. More

Neural networks have been powering breakthroughs in artificial intelligence, including the large language models that are now being used in a wide range of applications, from finance, to human resources to healthcare. But these networks remain a black box whose inner workings engineers and scientists struggle to understand. Now, a team led by data and computer scientists at the University of California San Diego has given neural networks the equivalent of an X-ray to uncover how they actually learn.
The researchers found that a formula used in statistical analysis provides a streamlined mathematical description of how neural networks, such as GPT-2, a precursor to ChatGPT, learn relevant patterns in data, known as features. This formula also explains how neural networks use these relevant patterns to make predictions.
“We are trying to understand neural networks from first principles,” said Daniel Beaglehole, a Ph.D. student in the UC San Diego Department of Computer Science and Engineering and co-first author of the study. “With our formula, one can simply interpret which features the network is using to make predictions.”
The team presented their findings in the March 7 issue of the journal Science.
Why does this matter? AI-powered tools are now pervasive in everyday life. Banks use them to approve loans. Hospitals use them to analyze medical data, such as X-rays and MRIs. Companies use them to screen job applicants. But it’s currently difficult to understand the mechanism neural networks use to make decisions and the biases in the training data that might impact this.
“If you don’t understand how neural networks learn, it’s very hard to establish whether neural networks produce reliable, accurate, and appropriate responses,” said Mikhail Belkin, the paper’s corresponding author and a professor at the UC San Diego Halicioglu Data Science Institute. “This is particularly significant given the rapid recent growth of machine learning and neural net technology.”
The study is part of a larger effort in Belkin’s research group to develop a mathematical theory that explains how neural networks work. “Technology has outpaced theory by a huge amount,” he said. “We need to catch up.”
The team also showed that the statistical formula they used to understand how neural networks learn, known as Average Gradient Outer Product (AGOP), could be applied to improve performance and efficiency in other types of machine learning architectures that do not include neural networks.

“If we understand the underlying mechanisms that drive neural networks, we should be able to build machine learning models that are simpler, more efficient and more interpretable,” Belkin said. “We hope this will help democratize AI.”
The machine learning systems that Belkin envisions would need less computational power, and therefore less power from the grid, to function. These systems also would be less complex and so easier to understand.
Illustrating the new findings with an example
(Artificial) neural networks are computational tools to learn relationships between data characteristics (i.e. identifying specific objects or faces in an image). One example of a task is determining whether in a new image a person is wearing glasses or not. Machine learning approaches this problem by providing the neural network many example (training) images labeled as images of “a person wearing glasses” or “a person not wearing glasses.” The neural network learns the relationship between images and their labels, and extracts data patterns, or features, that it needs to focus on to make a determination. One of the reasons AI systems are considered a black box is because it is often difficult to describe mathematically what criteria the systems are actually using to make their predictions, including potential biases. The new work provides a simple mathematical explanation for how the systems are learning these features.
Features are relevant patterns in the data. In the example above, there are a wide range of features that the neural networks learns, and then uses, to determine if in fact a person in a photograph is wearing glasses or not. One feature it would need to pay attention to for this task is the upper part of the face. Other features could be the eye or the nose area where glasses often rest. The network selectively pays attention to the features that it learns are relevant and then discards the other parts of the image, such as the lower part of the face, the hair and so on.
Feature learning is the ability to recognize relevant patterns in data and then use those patterns to make predictions. In the glasses example, the network learns to pay attention to the upper part of the face. In the new Science paper, the researchers identified a statistical formula that describes how the neural networks are learning features.

Alternative neural network architectures: The researchers went on to show that inserting this formula into computing systems that do not rely on neural networks allowed these systems to learn faster and more efficiently.
“How do I ignore what’s not necessary? Humans are good at this,” said Belkin. “Machines are doing the same thing. Large Language Models, for example, are implementing this ‘selective paying attention’ and we haven’t known how they do it. In our Science paper, we present a mechanism explaining at least some of how the neural nets are ‘selectively paying attention.'”
Study funders included the National Science Foundation and the Simons Foundation for the Collaboration on the Theoretical Foundations of Deep Learning. Belkin is part of NSF-funded and UC San Diego-led The Institute for Learning-enabled Optimization at Scale, or TILOS. More