Computers Math

An Oregon State University doctoral student and researchers at Adobe have created a new, cost-effective training technique for artificial intelligence systems that aims to make them less socially biased.
Eric Slyman of the OSU College of Engineering and the Adobe researchers call the novel method FairDeDup, an abbreviation for fair deduplication. Deduplication means removing redundant information from the data used to train AI systems, which lowers the high computing costs of the training.
Datasets gleaned from the internet often contain biases present in society, the researchers said. When those biases are codified in trained AI models, they can serve to perpetuate unfair ideas and behavior.
By understanding how deduplication affects bias prevalence, it’s possible to mitigate negative effects — such as an AI system automatically serving up only photos of white men if asked to show a picture of a CEO, doctor, etc. when the intended use case is to show diverse representations of people.
“We named it FairDeDup as a play on words for an earlier cost-effective method, SemDeDup, which we improved upon by incorporating fairness considerations,” Slyman said. “While prior work has shown that removing this redundant data can enable accurate AI training with fewer resources, we find that this process can also exacerbate the harmful social biases AI often learns.”
Slyman presented the FairDeDup algorithm last week in Seattle at the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
FairDeDup works by thinning the datasets of image captions collected from the web through a process known as pruning. Pruning refers to choosing a subset of the data that’s representative of the whole dataset, and if done in a content-aware manner, pruning allows for informed decisions about which parts of the data stay and which go.

“FairDeDup removes redundant data while incorporating controllable, human-defined dimensions of diversity to mitigate biases,” Slyman said. “Our approach enables AI training that is not only cost-effective and accurate but also more fair.”
In addition to occupation, race and gender, other biases perpetuated during training can include those related to age, geography and culture.
“By addressing biases during dataset pruning, we can create AI systems that are more socially just,” Slyman said. “Our work doesn’t force AI into following our own prescribed notion of fairness but rather creates a pathway to nudge AI to act fairly when contextualized within some settings and user bases in which it’s deployed. We let people define what is fair in their setting instead of the internet or other large-scale datasets deciding that.”
Collaborating with Slyman were Stefan Lee, an assistant professor in the OSU College of Engineering, and Scott Cohen and Kushal Kafle of Adobe. More

In a study published in Nature Communications June 19, a team of scientists led by Rice University’s Qimiao Si predicts the existence of flat electronic bands at the Fermi level, a finding that could enable new forms of quantum computing and electronic devices.
Quantum materials are governed by the rules of quantum mechanics, where electrons occupy unique energy states. These states form a ladder with the highest rung called the Fermi energy.
Electrons, being charged, repel each other and move in correlated ways. Si’s team found that electron interactions can create new flat bands at the Fermi level, enhancing their importance.
“Most flat bands are located far from the Fermi energy, which limits their impact on the material’s properties,” said Si, the Harry C. and Olga K. Wiess Professor of Physics and Astronomy at Rice.
Typically, a particle’s energy changes with its momentum. But in quantum mechanics, electrons can exhibit quantum interference, where their energy remains flat even when their momentum changes. These are known as flat bands.
“Flat electronic bands can enhance electron interactions, potentially creating new quantum phases and unusual low-energy behaviors,” Si said.
These bands are especially sought after in transition metal ions called d-electron materials with specific crystal lattices, where they often show unique properties, Si said.

The team’s findings suggest new ways to design these, which could inspire new applications for these materials in quantum bits, qubits and spintronics. Their research shows that electron interactions can link immobile and mobile electron states.
Using a theoretical model, the researchers demonstrated that these interactions can create a new type of Kondo effect, where immobile particles gain mobility by interacting with mobile electrons at the Fermi energy. The Kondo effect describes the scattering of conduction electrons in a metal due to magnetic impurities, resulting in a characteristic change in electrical resistivity with temperature.
“Quantum interference can enable the Kondo effect, allowing us to make significant progress,” said Lei Chen, a Ph.D. student at Rice.
A key attribute of the flat bands is their topology, Chen said. “The flat bands pinned to the Fermi energy provide a means to realize new quantum states of matter,” he said.
The team’s research reveals that this includes anyons and Weyl fermions, or massless quasiparticles and fermions that carry an electric charge. The researchers found that anyons are promising agents for qubits, and materials that host Weyl fermions may find applications in spin-based electronics.
The study also highlights the potential for these materials to be very responsive to external signals and capable of advanced quantum control. The results indicate that the flat bands could lead to strongly correlated topological semimetals at relatively low temperatures potentially operating at high temperatures or even room temperature.
“Our work provides the theoretical foundation for utilizing flat bands in strongly interacting settings to design and control novel quantum materials that operate beyond the realm of low temperatures,” Si said.
Contributors to this research include Fang Xie and Shouvik Sur, Rice postdoctoral associates of physics and astronomy; Haoyu Hu, Rice alumnus and postdoctoral fellow at Donostia International Physics Center; Silke Paschen, physicist at the Vienna University of Technology; and Jennifer Cano, theoretical physicist at Stony Brook University and the Flatiron Institute. More

Computers have come so far in terms of their power and potential, rivaling and even eclipsing human brains in their ability to store and crunch data, make predictions and communicate. But there is one domain where human brains continue to dominate: energy efficiency.
“The most efficient computers are still approximately four orders of magnitude — that’s 10,000 times — higher in energy requirements compared to the human brain for specific tasks such as image processing and recognition, although they outperform the brain in tasks like mathematical calculations,” said UC Santa Barbara electrical and computer engineering Professor Kaustav Banerjee, a world expert in the realm of nanoelectronics. “Making computers more energy efficient is crucial because the worldwide energy consumption by on-chip electronics stands at #4 in the global rankings of nation-wise energy consumption, and it is increasing exponentially each year, fueled by applications such as artificial intelligence.” Additionally, he said, the problem of energy inefficient computing is particularly pressing in the context of global warming, “highlighting the urgent need to develop more energy-efficient computing technologies.”
Neuromorphic (NM) computing has emerged as a promising way to bridge the energy efficiency gap. By mimicking the structure and operations of the human brain, where processing occurs in parallel across an array of low power-consuming neurons, it may be possible to approach brain-like energy efficiency. In a paper published in thejournal Nature Communications, Banerjee and co-workers Arnab Pal, Zichun Chai, Junkai Jiang and Wei Cao, in collaboration with researchers Vivek De and Mike Davies from Intel Labs propose such an ultra-energy efficient platform, using 2D transition metal dichalcogenide (TMD)-based tunnel-field-effect transistors (TFETs). Their platform, the researchers say, can bring the energy requirements to within two orders of magnitude (about 100 times) with respect to the human brain.
Leakage currents and subthreshold swing
The concept of neuromorphic computing has been around for decades, though the research around it has intensified only relatively recently. Advances in circuitry that enable smaller, denser arrays of transistors, and therefore more processing and functionality for less power consumption are just scratching the surface of what can be done to enable brain-inspired computing. Add to that an appetite generated by its many potential applications, such as AI and the Internet-of-Things, and it’s clear that expanding the options for a hardware platform for neuromorphic computing must be addressed in order to move forward.
Enter the team’s 2D tunnel-transistors. Emerging out of Banerjee’s longstandingresearch efforts to develop high-performance, low-power consumption transistors to meet the growing hunger for processing without a matching increase in power requirement, these atomically thin, nanoscale transistors are responsive at low voltages, and as the foundation of the researchers’ NM platform, can mimic the highly energy efficient operations of the human brain. In addition to lower off-state currents, the 2D TFETs also have a low subthreshold swing (SS), a parameter that describes how effectively a transistor can switch from off to on. According to Banerjee, a lower SS means a lower operating voltage, and faster and more efficient switching.
“Neuromorphic computing architectures are designed to operate with very sparse firing circuits,” said lead author Arnab Pal, “meaning they mimic how neurons in the brain fire only when necessary.” In contrast to the more conventional von Neumann architecture of today’s computers, in which data is processed sequentially, memory and processing components are separated and which continuously draw power throughout the entire operation, an event-driven system such as a NM computer fires up only when there is input to process, and memory and processing are distributed across an array of transistors. Companies like Intel and IBM have developed brain-inspired platforms, deploying billions of interconnected transistors and generating significant energy savings.

However, there’s still room for energy efficiency improvement, according to the researchers.
“In these systems, most of the energy is lost through leakage currents when the transistors are off, rather than during their active state,” Banerjee explained. A ubiquitous phenomenon in the world of electronics, leakage currents are small amounts of electricity that flow through a circuit even when it is in the off state (but still connected to power). According to the paper, current NM chips use traditional metal-oxide-semiconductor field-effect transistors (MOSFETs) which have a high on-state current, but also high off-state leakage. “Since the power efficiency of these chips is constrained by the off-state leakage, our approach — using tunneling transistors with much lower off-state current — can greatly improve power efficiency,” Banerjee said.
When integrated into a neuromorphic circuit, which emulates the firing and reset of neurons, the TFETs proved themselves more energy efficient than state-of-the-art MOSFETs, particularly the FinFETs (a MOSFET design that incorporates vertical “fins” as a way to provide better control of switching and leakage). TFETs are still in the experimental stage, however the performance and energy efficiency of neuromorphic circuits based on them makes them a promising candidate for the next generation of brain-inspired computing.
According to co-authors Vivek De (Intel Fellow) and Mike Davies (Director of Intel’s Neuromorphic Computing Lab), “Once realized, this platform can bring the energy consumption in chips to within two orders of magnitude with respect to the human brain — not accounting for the interface circuitry and memory storage elements. This represents a significant improvement from what is achievable today.”
Eventually, one can realize three-dimensional versions of these 2D-TFET based neuromorphic circuits to provide even closer emulation of the human brain, added Banerjee, widely recognized as one of the key visionaries behind 3D integrated circuits that are now witnessing wide scale commercial proliferation. More

Researchers have found a way to bind engineered skin tissue to the complex forms of humanoid robots. This brings with it potential benefits to robotic platforms such as increased mobility, self-healing abilities, embedded sensing capabilities and an increasingly lifelike appearance. Taking inspiration from human skin ligaments, the team, led by Professor Shoji Takeuchi of the University of Tokyo, included special perforations in a robot face, which helped a layer of skin take hold. Their research could be useful in the cosmetics industry and to help train plastic surgeons.
Takeuchi is a pioneer in the field of biohybrid robotics, where biology and mechanical engineering meet. So far, his lab, the Biohybrid Systems Laboratory, has created mini robots that walk using biological muscle tissue, 3D printed lab-grown meat, engineered skin that can heal, and more. It was during research on the last of these items that Takeuchi felt the need to take the idea of robotic skin further to improve its properties and capabilities.
“During previous research on a finger-shaped robot covered in engineered skin tissue we grew in our lab, I felt the need for better adhesion between the robotic features and the subcutaneous structure of the skin,” said Takeuchi. “By mimicking human skin-ligament structures and by using specially made V-shaped perforations in solid materials, we found a way to bind skin to complex structures. The natural flexibility of the skin and the strong method of adhesion mean the skin can move with the mechanical components of the robot without tearing or peeling away.”
Previous methods to attach skin tissue to solid surfaces involved things like mini anchors or hooks, but these limited the kinds of surfaces that could receive skin coatings and could cause damage during motion. By carefully engineering small perforations instead, essentially any shape of surface can have skin applied to it. The trick the team employed was to use a special collagen gel for adhesion, which is naturally viscous so difficult to feed into the minuscule perforations. But using a common technique for plastic adhesion called plasma treatment, they managed to coax the collagen into the fine structures of the perforations while also holding the skin close to the surface in question.
“Manipulating soft, wet biological tissues during the development process is much harder than people outside the field might think. For instance, if sterility is not maintained, bacteria can enter and the tissue will die,” said Takeuchi. “However, now that we can do this, living skin can bring a range of new abilities to robots. Self-healing is a big deal — some chemical-based materials can be made to heal themselves, but they require triggers such as heat, pressure or other signals, and they also do not proliferate like cells. Biological skin repairs minor lacerations as ours does, and nerves and other skin organs can be added for use in sensing and so on.”
This research was not just made to prove a point, though. Takeuchi and his lab have a goal in mind for this application that could help in several areas of medical research. The idea of an organ-on-a-chip is not especially new, and finds use in things like drug development, but something like a face-on-a-chip could be useful in research into skin aging, cosmetics, surgical procedures, plastic surgery and more. Also, if sensors can be embedded, robots may be endowed with better environmental awareness and improved interactive capabilities.
“In this study, we managed to replicate human appearance to some extent by creating a face with the same surface material and structure as humans,” said Takeuchi. “Additionally, through this research, we identified new challenges, such as the necessity for surface wrinkles and a thicker epidermis to achieve a more humanlike appearance. We believe that creating a thicker and more realistic skin can be achieved by incorporating sweat glands, sebaceous glands, pores, blood vessels, fat and nerves. Of course, movement is also a crucial factor, not just the material, so another important challenge is creating humanlike expressions by integrating sophisticated actuators, or muscles, inside the robot. Creating robots that can heal themselves, sense their environment more accurately and perform tasks with humanlike dexterity is incredibly motivating.” More

Every so often, a food product is recalled because of some sort of contamination. For consumers of such products, a recall can trigger doubt in the safety and reliability of what they eat and drink. In many cases, a recall will come too late to keep some people from getting ill.
In spite of the food industry’s efforts to fight pathogens, products are still contaminated and people still get sick. Much of the problem stems from the tools available to screen for harmful pathogens, which are often not effective enough at protecting the public.
In AIP Advances, by AIP Publishing, researchers from Guangdong University of Technology and Pudong New District People’s Hospital developed a new method for detecting foodborne pathogens that is faster, cheaper, and more effective than existing methods. The researchers hope their technique can improve screening processes and keep contaminated food out of the hands of consumers.
Even with the best detection method, finding contaminating pathogens is not an easy task.
“Detecting these pathogens is challenging, due to their diverse nature and the various environments in which they can thrive,” said author Silu Feng. “Additionally, low concentrations of pathogens in large food samples, the presence of similar non-pathogenic organisms, and the complex nature of different food types make accurate and rapid detection difficult.”
Existing detection methods do exist, such as cell culture and DNA sequencing, but are challenging to employ at large scales. Not every batch of food can be thoroughly tested, so some contaminants inevitably slip through.
“Overall, these methods face limitations such as lengthy result times, the need for specialized equipment and trained personnel, and challenges in detecting multiple pathogens simultaneously, highlighting the need for improved detection techniques,” said Feng.

The study’s authors decided to take a different approach, designing a microfluidic chip that uses light to detect multiple types of pathogens simultaneously. Their chip is created using 3D printing, making it easy to fabricate in large amounts and modify to target specific pathogens.
The chip is split into four sections, each of which is tailored to detect a specific pathogen. If that pathogen is present in the sample, it will bind to a detection surface and change its optical properties. This arrangement let the researchers detect several common bacteria, such as E. coli, salmonella, listeria, and S. aureus, quickly and at very low concentrations.
“This method can quickly and effectively detect multiple different pathogens, and the detection results are easy to interpret, significantly improving detection efficiency,” said Feng.
The team plans to continue developing their device to make it even more applicable for food screening. More

Meet CARMEN, short for Cognitively Assistive Robot for Motivation and Neurorehabilitation-a small, tabletop robot designed to help people with mild cognitive impairment (MCI) learn skills to improve memory, attention, and executive functioning at home.
Unlike other robots in this space, CARMEN was developed by the research team at the University of California San Diego in collaboration with clinicians, people with MCI, and their care partners. To the best of the researchers’ knowledge, CARMEN is also the only robot that teaches compensatory cognitive strategies to help improve memory and executive function.
“We wanted to make sure we were providing meaningful and practical inventions,” said Laurel Riek, a professor of computer science and emergency medicine at UC San Diego and the work’s senior author.
MCI is an in-between stage between typical aging and dementia. It affects various areas of cognitive functioning, including memory, attention, and executive functioning. About 20% of individuals over 65 have the condition, with up to 15% transitioning to dementia each year. Existing pharmacological treatments have not been able to slow or prevent this evolution, but behavioral treatments can help.
Researchers programmed CARMEN to deliver a series of simple cognitive training exercises. For example, the robot can teach participants to create routine places to leave important objects, such as keys; or learn note taking strategies to remember important things. CARMEN does this through interactive games and activities.
The research team designed CARMEN with a clear set of criteria in mind. It is important that people can use the robot independently, without clinician or researcher supervision. For this reason, CARMEN had to be plug and play, without many moving parts that require maintenance. The robot also has to be able to function with limited access to the internet, as many people do not have access to reliable connectivity. CARMEN needs to be able to function over a long period of time. The robot also has to be able to communicate clearly with users; express compassion and empathy for a person’s situation; and provide breaks after challenging tasks to help sustain engagement.
Researchers deployed CARMEN for a week in the homes of several people with MCI, who then engaged in multiple tasks with the robot, such as identifying routine places to leave household items so they don’t get lost, and placing tasks on a calendar so they won’t be forgotten. Researchers also deployed the robot in the homes of several clinicians with experience working with people with MCI. Both groups of participants completed questionnaires and interviews before and after the week-long deployments.

After the week with CARMEN, participants with MCI reported trying strategies and behaviors that they previously had written off as impossible. All participants reported that using the robot was easy. Two out of the three participants found the activities easy to understand, but one of the users struggled. All said they wanted more interaction with the robot.
“We found that CARMEN gave participants confidence to use cognitive strategies in their everyday life, and participants saw opportunities for CARMEN to exhibit greater levels of autonomy or be used for other applications,” the researchers write.
The research team presented their findings at the ACM/IEEE Human Robot Interaction (HRI) conference in March 2024, where they received a best paper award nomination.
Next steps
Next steps include deploying the robot in a larger number of homes.
Researchers also plan to give CARMEN the ability to have conversations with users, with an emphasis on preserving privacy when these conversations happen. This is both an accessibility issue (as some users might not have the fine motor skills necessary to interact with CARMEN’s touch screen), as well as because most people expect to be able to have conversations with systems in their homes. At the same time, researchers want to limit how much information CARMEN can give users. “We want to be mindful that the user still needs to do the bulk of the work, so the robot can only assist and not give too many hints,” Riek said.
Researchers are also exploring how CARMEN could assist users with other conditions, such as ADHD.
The UC San Diego team built CARMEN based on the FLEXI robot from the University of Washington. But they made substantial changes to its hardware, and wrote all its software from scratch. Researchers used ROS for the robot’s operating system.
Many elements of the project are available at https://github.com/UCSD-RHC-Lab/CARMEN More

A novel technique with potential applications for fields such as droplet chemistry and photochemistry has been demonstrated by an Osaka Metropolitan University-led research group.
Professor Yasuyuki Tsuboi of the Graduate School of Science and the team investigated Förster resonance energy transfer (FRET), a phenomenon seen in photosynthesis and other natural processes where a donor molecule in an excited state transfers energy to an acceptor molecule.
Using dyes to mark the donor and acceptor molecules, the team set out to see if FRET could be controlled by the intensity of an optical force, in this case a laser beam. By focusing a laser beam on an isolated polymer droplet, the team showed that increased intensity accelerated the energy transfer, made visible by the polymer changing color due to the dyes mixing.
Fluorescence could also be controlled just by adjusting the laser intensity without touching the sample, offering a novel non-contact approach.
“Although this research is still at a basic stage, it may provide new options for a variety of future FRET research applications,” Professor Tsuboi explained. “We believe that extending this to quantum dots as well as new polymer systems and fluorescent molecules is the next challenge.” More

Artificial intelligence (AI) is accelerating the development of new materials. A prerequisite for AI in materials research is large-scale use and exchange of data on materials, which is facilitated by a broad international standard. A major international collaboration now presents an extended version of the OPTIMADE standard.
New technologies in areas such as energy and sustainability involving for example batteries, solar cells, LED lighting and biodegradable materials require new materials. Many researchers around the world are working to create materials that have not existed before. But there are major challenges in creating materials with the exact properties required, such as not containing environmentally hazardous substances and at the same time being durable enough to not break down.
“We’re now seeing an explosive development where researchers in materials science are adopting AI methods from other fields and also developing their own models to use in materials research. Using AI to predict properties of different materials opens up completely new possibilities,” says Rickard Armiento, associate professor at the Department of Physics, Chemistry and Biology (IFM) at Linköping University in Sweden.
Today, many demanding simulations are performed on supercomputers that describe how electrons move in materials, which gives rise to different material properties. These advanced calculations yield large amounts of data that can be used to train machine learning models.
These AI models can then immediately predict the responses to new calculations that have not yet been made, and by extension predict the properties of new materials. But huge amounts of data are required to train the models.
“We’re moving into an era where we want to train models on all data that exist,” says Rickard Armiento.
Data from large-scale simulations, and general data about materials, are collected in large databases. Over time, many such databases have emerged from different research groups and projects, like isolated islands in the sea. They work differently and often use properties that are defined in different ways.
“Researchers at universities or in industry who want to map materials on a large scale or want to train an AI model must retrieve information from these databases. Therefore, a standard is needed so that users can communicate with all these data libraries and understand the information they receive,” says Gian-Marco Rignanese, professor at the Institute of Condensed Matter and Nanosciences at UCLouvain in Belgium.
The OPTIMADE (Open databases integration for materials design) standard has been developed over the past eight years. Behind this standard is a large international network with over 30 institutions worldwide and large materials databases in Europe and the USA. The aim is to give users easier access to both leading and lesser-known materials databases. A new version of the standard, v1.2, is now being released, and is described in an article published in the journal Digital Discovery. One of the biggest changes in the new version is a greatly enhanced possibility to accurately describe different material properties and other data using common, well-founded definitions.
The international collaboration spans the EU, the UK, the US, Mexico, Japan and China together with institutions such as École Polytechnique Fédérale de Lausanne (EPFL), University of California Berkeley, University of Cambridge, Northwestern University, Duke University, Paul Scherrer Institut, and Johns Hopkins University. Much of the collaboration takes place in meetings with annual workshops funded by CECAM (Centre Européen de Calcul Atomique et Moléculaire) in Switzerland, with the first one funded by the Lorentz Center in the Netherlands. Other activities have been supported by the organisation Psi-k, the competence centre NCCR MARVEL in Switzerland, and the e-Science Research Centre (SeRC) in Sweden. The researchers in the collaboration receive support from many different financiers. More