Articles tagged with Neural Networks
A novel software and hardware solution analyzes synaptic dysfunctions related to Autism Spectrum Disorders and similar complex neurological conditions. The technology, comprising a neural network model and hardware architecture, provides real-time, biologically relevant synaptic analysis.
Scientists have discovered that when brain regions agree, their shared activity lasts longer, but when they disagree, the mismatch quickly fades. This consensus-building mechanism could lead to a deeper understanding of brain function and inspire new AI system designs.
Researchers found that the amygdalostriatal transition zone, a small brain region near the amygdala, is responsible for sustained fear responses. The study provides new insights into the brain's fear circuitry and may lead to the development of more effective therapies for fear and panic disorders.
Researchers developed a neural network-based framework that significantly improves the prediction of satellite clock bias for low earth orbit satellites. The approach uses advanced data-processing strategies and a reconstruction fine-tuning mechanism to enable more accurate and stable real-time predictions of satellite timing errors.
Researchers have developed flexible and ultra-fast artificial synapses printed entirely from room-temperature liquid inks. These brain-inspired chips can process health data directly on the body and dissolve when no longer needed, eliminating the need for extreme vacuum chambers and rare metals.
Researchers developed an AI-guided laser technique to carve micro-pyramids for robots to sense soft surfaces gently. The technique enables the creation of flexible conductive skins with high sensitivity and linearity, outperforming conventional designs.
A new learning mechanism uses natural variability in neural activity to understand how synapses adapt and improve the learning capabilities of brain-inspired devices. The mechanism, called Spike-based Alignment Learning, solves the weight transport problem and matches the performance of existing approaches without unrealistic assumptions.
A new study used neural networks to reveal how different kinds of training can change how learning happens. Training a simpler task first made neural networks respond more accurately to complex tasks, mirroring how living brains learn.
Researchers propose an AI framework, called interoceptive AI, that uses internal states to inform learning and decision-making in dynamic environments. This approach treats internal conditions as a continuous source of context, influencing what an agent learns, prioritizes, and does.
Optical convolution computation enables parallel light propagation and multiplexing for faster and more energy-efficient computing systems. The review organizes the field into two paradigms: definition-based and theorem-based, which leverage mathematical principles to implement convolution operations in the physical domain.
Researchers develop a novel framework, LL-Refiner, to enhance high-resolution images in poor lighting conditions, outperforming state-of-the-art techniques. The framework uses a coarse enhancement stage to guide the recovery of fine details, resulting in improved visual quality and performance in downstream computer-vision tasks.
Four assistant professors, Yahong Yang, Sammy Luo, Lebing Chen, and Kunyan Zhang, join Binghamton University as Simons Empire Faculty Fellows, bringing expertise in quantum materials and artificial intelligence. Their research focuses on developing new technologies, including energy-efficient systems and next-generation sensing platforms.
A new study suggests that specific brain waves during sleep provide protection against the cognitive decline associated with elevated levels of the neurotransmitter orexin. Individuals with stronger sleep spindles and sleep oscillations during nonrapid eye movement (NREM) sleep had less cognitive decline over time.
A novel AI model has been developed that can recognize yoga poses with high accuracy, paving the way for more effective digital coaching tools and movement-monitoring applications. The model achieved accuracy levels of over 93% during testing, significantly outperforming previous models.
A study by Bar-Ilan University researchers found that learning is driven primarily by changes in the strength of existing neural connections. The models became significantly better at learning as the amount of training data increased, but the proportion of lost connections remained roughly the same.
Researchers have developed a novel impedance-gradient metadevice that bridges structural engineering and AI for next-generation stealth systems. The device achieves ultra-broadband microwave absorption spanning the full 2–18 GHz radar band while simultaneously achieving infrared thermal insulation and rapid visible color adaptation.
Researchers have developed a new version of the Daydreaming algorithm, which combines learning and cleaning to improve artificial memory systems' reliability even with biased data. The algorithm focuses on differences between pixels, allowing it to work effectively with strongly biased data, similar to real-world conditions.
The Data Sciences Institute at the University of Toronto has been awarded $1 million in Claude API credits to support AI-enabled research. Researchers will gain access to cutting-edge AI tools, enabling discovery, analysis, and innovation across disciplines.
A new study finds that human cortical neurons have remarkable computational capabilities, surpassing those of other mammals. The researchers developed a new method to measure the complexity of individual neurons, revealing their sophisticated computing power.
Researchers found that local electric fields exert influence on neurons via ephaptic coupling, which helps explain variations in brain activity even within the same task. The study suggests that manipulating these electric fields could be a potential therapeutic approach for improving brain function in disease.
Researchers developed an on-chip all-optical supernode for ultra-low-latency deep neural network inference, achieving a 100-fold increase in inference speed while using only one-ninth of computing resources. The system supports high-speed data routing and switching with low loss and flat response over a spectral range exceeding 100 nm.
Scientists have discovered a new type of brain wave that rotates over space and time, relying on a circular anatomical circuit in the sensory cortex. This rotation is coordinated between different brain regions, including sensory and motor parts, and may play a role in sharing information across these areas.
A neural network-based machine learning model accelerates diffuse optical tomography by over a million-fold, enabling real-time diagnosis. The model accurately reproduces signals even for unseen parameter combinations, with each inference taking approximately 2 milliseconds.
A digital 'super-brain' with physics-based knowledge significantly speeds up the design and development of optical components, such as those for quantum computers and camera lenses. By integrating physical principles into machine learning algorithms, researchers reduce simulation time from months to days.
Researchers have developed a wearable sensor that reads chemical signatures of human breath to decode silent speech into text. The device uses a microscopic nanoforest to capture rapid water vapor changes, achieving 98.51% accuracy rate.
A new study published in the Journal of Big Data highlights the journal's emergence as a leading publication in data science and artificial intelligence research. The study found that JBD has become a central hub for high-impact research worldwide, with significant contributions from top researchers.
Researchers developed a neural network approach that learns to clean co-movement patterns in markets before building portfolios. The method achieved lower volatility and higher Sharpe ratios compared to traditional methods.
TEGNet accelerates optimization in thermoelectric generator design by predicting performance with high accuracy and speed. The AI model enables designers to freely combine independent models for various materials, enabling complex structure exploration and high conversion efficiencies.
Researchers have developed a simplified mathematical model of learning in neural networks, shedding new light on how these systems produce their responses. The toy model, inspired by physics principles, captures key features of complex systems and offers insights into the surprising efficiency and stability of modern AI systems.
HelixAI develops AI-driven platform for researchers and clinicians to integrate complex biomedical data, improving diagnosis and prognosis. It also launches Helix for Longevity, a consumer-facing application estimating biological age and providing personalized recommendations for health promotion.
Artificial synapses are built from soft, bio-friendly materials that operate like human brain synapses, merging data storage and computing into a single unit. Laboratory prototypes demonstrate immense capabilities, consuming energy on the scale of femtojoules.
Researchers at Tohoku University demonstrated that living biological neurons can be trained to perform a supervised temporal pattern learning task. The study integrates cultured neuronal networks into a machine learning framework, generating complex time-series signals comparable to those involved in motor control.
Researchers developed two neural network models predicting human biological age based on blood biochemistry and gut microbiota. The models demonstrated high predictive accuracy and explainability, holding potential for monitoring intervention effects in clinical trials.
Researchers identified new neurons that respond to different spatial frequencies, allowing for more precise object recognition, and used digital twins to confirm the findings in mouse brains.
Researchers developed photonic computing chips that enable fast, all-optical learning and decision making, overcoming key limitations for photonic spiking neural systems. The new chips could improve autonomous driving technologies and enable robotic systems that learn through real-world interactions.
Researchers developed a new method called Learn-to-Steer, which analyzes internal attention patterns of image-generation models to guide their placement according to user instructions. The approach improved accuracy in understanding spatial relationships by up to 61% in existing trained models.
Researchers found that dosed nonlinearity improves model performance in various tasks, especially with limited data. Nonlinear units function like flexible switches, adapting linear processing modes based on context.
Researchers created a new frequency-aware approach to crafting adversarial images that better match human visual perception. The method, called Input-Frequency Adaptive Adversarial Perturbation (IFAP), significantly outperformed existing techniques in structural and textural similarity.
Engineers at the University of Pennsylvania have discovered that foams exhibit internal motion resembling deep learning in AI systems. The study suggests a common mathematical principle underlying both foams and AI training, with implications for designing adaptive materials and understanding biological structures.
Researchers at Rockefeller University have made a breakthrough in understanding how the brain controls facial expressions, discovering a complex network of neural circuits involved. Contrary to long-held assumptions, both lower-level and higher-level brain regions are involved in encoding different types of facial gestures.
Researchers reveal that harm causes stronger guilt, while sense of responsibility triggers shame, influencing compensatory behaviors. The study also identifies distinct neural activity for guilt and shame-driven decisions, shedding light on the cognitive processes guiding these emotions.
This study applies Physics-Informed Neural Networks (PINNs) and Extreme Learning Machines to solve complex option pricing problems under stochastic volatility. The research enables accurate pricing of American-style options for both equity and real estate index derivatives, addressing a significant challenge in quantitative finance.
A research team developed an AI-guided framework to discover new metallic glasses by combining element embeddings learned from Wikipedia with graph neural networks. This approach overcomes challenges in predicting glass-forming systems, enabling the discovery of promising compositions with high glass-forming ability.
Researchers at Institute of Science Tokyo developed a neural-network-based 3D imaging technique that can precisely measure moving objects. The new method reconstructs high-resolution 3D shapes using only three projection patterns, enabling dynamic 3D measurement across various applications.
KVzip reduces chatbot response time and memory cost while maintaining accuracy, achieving 3–4× memory reduction and approximately 2× faster response times. The technology also demonstrates scalability to extremely long contexts and has been integrated into NVIDIA's open-source library.
Autograph, a new framework, uses graph neural networks and deep reinforcement learning to achieve higher accuracy and faster execution of compute-intensive programs. It outperformed other approaches across various datasets, with notable improvements on Polybench, NPB, and SPEC 2006 benchmarks.
Recent advances in deep learning techniques have overcome limitations in spatial and temporal resolution of BOLD-fMRI. DL models improve image quality through super-resolution reconstruction, automate segmentation, and enhance registration, enabling finer localization of neural activity and more precise brain activity quantification.
Researchers developed a machine learning-based workflow, SPaDe-CSP, to predict crystal structures of organic molecules. The workflow narrows the search space by predicting probable space groups and crystal densities before computationally intensive relaxation steps.
A team of researchers at the University of Waterloo developed a framework that uses mathematical tools and machine learning to rigorously check and verify the safety of AI-driven systems. The framework has been tested on challenging control problems and matched or exceeded traditional approaches.
A team of researchers from Yokohama National University has developed a novel compact superconductive neuron device that operates at high speeds with ultra-low power consumption. The device eliminates variation in elemental circuit characteristics, achieving ideal input-output characteristics and resolving the vanishing gradient problem.
VFF-Net applies label-wise noise labelling, cosine similarity-based contrastive loss, and layer grouping to improve image classification performance compared to conventional forward-forward networks. The algorithm reduces test errors on various datasets, enabling lighter and more brain-like training methods that make AI more sustainable.
Salk scientists pinpoint gracile nucleus as brain area responsible for differentiating between painful and non-painful touch, with dysfunction leading to chronic pain. Altered neuronal activity in the dorsal column nuclei drives mechanical allodynia, causing the brain to misinterpret innocuous light touch as painful.
Researchers have identified a new population of hypothalamic neurons, Crabp1 neurons, that play a critical role in regulating energy expenditure. Silencing these neurons leads to reduced energy expenditure and obesity, while activating them enhances locomotor activity and protects against high-fat diet-induced weight gain.
Researchers have developed an emulator called Effort.jl that mimics the behavior of large-scale structure models, allowing for fast analysis on standard laptops. The new model delivers similar accuracy as the original, enabling scientists to analyze upcoming data releases from experiments like DESI and Euclid.
The team will study neurons within a brain organoid, a millimeter-sized, three-dimensional structure grown in the lab from adult stem cells, to design smarter and more sustainable artificial intelligence. They aim to replicate complex computations that occur in the human brain to improve AI efficiency.
Researchers at Tohoku University have developed an AI-built materials map that combines experimental data with computational predictions to identify promising materials for thermoelectric waste-heat recovery. The map enables faster development timelines and reduces trial-and-error, accelerating innovation in energy-related technologies.
Researchers developed a simple model that reproduces deep neural network features, allowing for optimized parameter tuning. The 'folding ruler' model demonstrates how nonlinearity and noise improve network performance, enabling more efficient training without trial-and-error.
A new photonic neural network developed in China achieves higher classification accuracy than digital models by using physical light transformations and multisynaptic optical paths. The system's design avoids errors introduced by translating software to hardware, marking a major step forward in optical AI hardware.
Researchers used Bayesian neural network to identify relationships between gut bacteria and metabolites, providing clues about health. The approach outperformed existing methods in analyzing sleep disorder, obesity, and cancer studies.
Researchers have successfully controlled a dexterous robotic hand using noninvasive EEG-based Brain-Computer Interfaces (BCIs) for individual finger movements. The study demonstrates real-time brain decoding and motor imagery control, paving the way for potential applications beyond basic communication to intricate motor control.