Articles tagged with Neural Networks
A team of Skoltech researchers demonstrates that universal adversarial perturbations (UAPs) can be explained by classical Turing patterns. This finding can help construct a theory of adversarial examples and design defenses against pattern recognition systems.
Scientists have found a way to reduce energy consumption in deep neural networks, paving the way for more efficient AI hardware. The approach uses simple electrical impulses instead of complex numerical values, maintaining high accuracy.
A new deep-learning algorithm, CARRL, is designed to help machines build a healthy skepticism of their measurements and inputs. By combining reinforcement-learning algorithms with deep neural networks, researchers created an approach that outperformed standard machine-learning techniques in scenarios with uncertain and adversarial inputs.
The new project posits that deep neural networks struggle with real-world problems due to an overemphasis on neurons, which neglect the role of astrocytes. Integrating astrocytes could enhance DNN efficiency and performance.
Researchers developed a new framework to analyze massive data from thousands of individual neurons, outperforming previous models. The method captures complex dynamics and fluctuations, offering insights into animal processing information and adapting to environmental changes.
An interdisciplinary team of biologists and computational researchers designed a neural network named BPNet that can interpret regulatory code by predicting transcription factor binding from DNA sequences with unprecedented accuracy. The model revealed novel insights, including a rule governing the binding of the well-studied transcrip...
AI researchers have developed a method to train neural networks to predict the function of DNA sequences, allowing for deciphering larger patterns. This breakthrough enables analysis of complex DNA sequences critical to development and disease, potentially improving understanding of gene regulation and its impact on diseases.
Researchers from RUDN University found a way to reduce the size of a trained neural network by six times without retraining, achieving significant storage volume reduction and minimal accuracy loss. The new method leverages correlations between initial and simplified weights, eliminating the need for post-training.
Researchers at Radboud University create a network of single atoms that mimic brain-like behavior and adapt to external stimuli. They plan to scale up the system and explore new materials to build self-learning computing devices.
Researchers developed a neural network that can adapt to new data inputs, continuously learning from changing time series data streams. This 'liquid' network could boost the development of emerging technologies like self-driving cars and medical diagnostics.
Researchers developed a neural network to assess sleep apnea severity in acute stroke patients, showing high accuracy and ease of use. The new screening method uses simple nocturnal pulse oximetry, enabling early detection and treatment of sleep apnea in cerebrovascular disease patients.
UT Arlington computer scientists develop a deep learning method to generate synthetic objects for robot training, overcoming the need for manual capture of images from human-centric perspectives. The technique uses generative adversarial networks (GANs) to create photorealistic full scenes and dense colored point clouds with fine details.
Researchers have developed a new optical neural network that can process large-scale data and images at incredible speeds, surpassing electronic computing hardware. This innovation has the potential to transform artificial intelligence in applications such as image recognition, medical diagnosis, and real-time video analysis.
Developing a new era of optical signal processing, researchers created an optical convolutional neural network accelerator capable of processing large amounts of information per second. This innovation harnesses the massive parallelism of light to outperform top-of-the-line graphics processing units by over one order of magnitude.
Researchers have proposed a new approach using neural networks and omics data to predict Z-DNA regions in the human genome. DeepZ, a recurrent neural network, achieved higher accuracy than existing algorithms, allowing for the mapping of potential Z-DNA sites across the entire genome.
Researchers analyzed over 6 million video clips from 144 countries to discover universal human emotional expression across cultures. The study found that people share about 70% of facial expressions in response to different social and emotional situations.
A team of researchers from Duke University has developed a method to make neural networks more transparent and interpretable. By modifying the reasoning process behind predictions, it is possible to better understand how these complex models work. The approach involves replacing standard parts of a neural network with new ones that con...
A study from Oregon Health & Science University found that damage to a small number of brain cells can stop activity across a vast network of neural circuits. This effect, known as the bystander effect, may help explain temporary but severe loss of cognitive function in traumatic brain injury or disease cases.
A new study published in EPJ B reveals how complex dynamics in branching networks of neurons can be predicted to trigger episodes of epilepsy. The team's findings could lead to the development of better early warning systems for patients.
A team of researchers has found that visual short-term memory retains multiple types of information, including color, texture, and name, in a single phase. This challenges previous assumptions about the complexity of visual short-term memory and highlights the importance of complex brain activity analysis.
A team from UNIGE used a specially developed video game to investigate the emergence of emotions, confirming that brain components respond in parallel distributed throughout the brain. The results show transient synchronisation generating an emotional state, involving areas like the somatosensory and motor pathways.
Researchers used connectomic mapping to study inhibitory neuronal circuitry in developing mice brains. They found that different types of interneurons followed distinct developmental time courses to establish synaptic partners.
Researchers have discovered subnetworks within BERT that can complete the same task more efficiently, reducing computing costs and increasing accessibility to state-of-the-art natural language processing. The 'lottery ticket hypothesis' identifies these leaner subnetworks, which can be repurposed for multiple tasks.
Researchers studied how convolutional neural networks respond to brightness and color visual illusions, finding that they are similarly deceived as humans. The study highlights the limitations of CNNs in mimicking human vision, revealing both similarities and differences between the two.
A neural network has been developed to estimate uncertainty, allowing for safer outcomes in AI-assisted decision-making. The 'deep evidential regression' approach accelerates uncertainty estimation, enabling faster and more accurate confidence levels, reducing the risk of errors.
Researchers at the University of California, Berkeley, have created AI software that gives robots speed and skill to grasp objects, making it feasible for them to assist humans in warehouses. The technology reduces computation time from 29 seconds to under one-tenth of a second.
Researchers compared neuronal networks to galaxy distributions, finding similarities in complexity and self-organization. The study suggests that diverse physical processes can create comparable structures despite vastly different scales.
A new system called MCUNet enables artificial intelligence on household appliances while improving data security and energy efficiency. The technology uses compact neural networks to deliver unprecedented speed and accuracy for deep learning on IoT devices.
Researchers at Penn State have developed graphene-based memory resistors that mimic the brain's neural networks and offer high precision neuromorphic computing. The new technology can control up to 16 possible memory states, compared to two in most existing memristors.
A new wrist-worn camera system enables accurate 3D hand pose estimation, outperforming previous work by 20%. The system's accuracy reaches 75% in detecting different grasp types and can be used for smart device control, virtual mice, and keyboards.
Researchers developed new AI models inspired by nature, reducing complexity and enhancing interpretability. These models can control vehicles with just a few artificial neurons, outperforming previous deep learning models in tasks such as autonomous lane keeping.
A new deep learning model inspired by tiny animals has shown decisive advantages over previous models in tasks such as autonomous driving. The model achieves better performance with fewer neurons and is more interpretable than complex 'black box' systems.
The study reveals how information flows between neuronal network clusters and how these clusters self-optimize over time. The findings can open new research directions for biologically inspired artificial intelligence, detection of brain cancer and diagnosis.
Researchers created an abstract language that describes protein molecules' shapes and structures, enabling predictions of their dynamics. This method uses machine learning algorithms to analyze molecular movements and provides insights into disease causes and targeted drug therapies.
Researchers are developing a novel deep learning technique to identify relationships between brain networks and Alzheimer's disease using algorithms mimicking neural networks. The goal is to pinpoint specific areas in the brain to slow and treat disease progression.
Scientists have created a new nanodevice that acts like a brain cell and can be joined to form networks that solve problems in a brain-like manner. These systems can identify possible mutations in a virus, relevant for ensuring vaccine efficacy.
A new data processing module called attentive normalization improves the performance of deep neural networks by combining feature normalization and feature attention. The hybrid module significantly increases accuracy while using negligible extra computational power, and facilitates better transfer learning between different domains.
A team of engineers and computer scientists are developing a theory of deep learning based on rigorous mathematical principles to improve reliability and predictability in AI systems. They will use three perspectives: local to global understanding, statistical analysis, and formal verification.
Researchers have found a way to improve the accuracy of brain-inspired computing systems using memristors, which are at least 1,000 times more energy-efficient than conventional transistor-based AI hardware. This could lead to a significant reduction in carbon emissions from training one AI model.
Researchers developed a neural network to assess stored blood quality, achieving 76.7% agreement with experts in identifying damaged red blood cells. The network outperformed expert predictions when trained using only storage duration.
Researchers create novel method using artificial intelligence to connect static and dynamic calculations, enhancing system security and safety requirements. The approach enables operators to anticipate disruptions and optimize resource allocation for a more resilient power grid.
Researchers found CAMSAP1 plays a crucial role in regulating axon/dendrite differentiation by creating an unbalanced distribution of microtubules among processes. The study resolves a long-standing question in neuroscience about the decisive factor for neuronal polarity establishment.
Researchers identified a new role for bi-directional connections in accelerating communication between brain regions. By creating loops, these connections can establish resonance and amplify signals, reducing the need for synchronization and increasing network efficiency.
Researchers are using multidisciplinary approaches, cutting-edge imaging technologies, and cyber resources to study synaptic weight and its effects on the brain. The team aims to determine what factors shape synaptic structures and function, shedding light on basic understanding of the brain.
Researchers at Tohoku University and the University of Gothenburg developed a novel voltage-controlled spintronic oscillator capable of closely imitating non-linear oscillatory neural networks. The technology allows for strong tuning with negligible energy consumption, enabling efficient training of large neural networks.
Scientists from Nanyang Technological University (NTU Singapore) have developed an AI system that recognizes human hand gestures with high precision. The system combines skin-like electronics with computer vision and achieves accuracy even in poor environmental conditions.
A new study by MIT neuroscientists provides a mathematical model showing how the brain overcomes unpredictable disturbances to produce reliable computations. The model describes an inclination toward robust stability built into neural circuits due to connections between neurons.
Researchers at the University of Illinois trained light-sensitive neurons using timed pulses of light during early cell development, leading to improved connections, responsivity, and gene expression. The early training resulted in long-lasting improvements, whereas cells trained later had transient responses.
In a breakthrough study, scientists successfully implanted highly specialized neural stem cell grafts directly into mouse spinal cord injuries, showing they integrated with host networks and behaved like neurons. The grafts displayed spontaneous activity, responded to sensory stimuli, and formed functional connections with host neurons.
A researcher at Sandia National Laboratories has won an Early Career Research Program award to develop methods for applying physics laws to observe large-scale physical events. The project aims to achieve a millionfold change in scale, from meter- to microscale features.
Researchers at Washington University in St. Louis developed a new algorithm called Parallel Residual Projection (PRP) to solve linear inverse problems by breaking them down into smaller tasks that can be solved in parallel on standard computers.
A team led by University of Pittsburgh's Jingtong Hu is working on a project to train algorithms that can accurately diagnose pneumonia caused by COVID-19 using CT scans. The goal is to create a mobile scanning device that can quickly screen for signs of the disease in crowded places.
A study by Heidelberg University and Max-Planck-Institute found that the distance to criticality can be adjusted in a brain-inspired chip, but only complex tasks benefit from it. Optimal network dynamics can be tuned using homeostatic plasticity by adapting mean input strength.
A groundbreaking study using MRI scans of 130 mammalian brains found that brain connectivity levels are equal in all species, including humans. The research revealed a universal law: Conservation of Brain Connectivity, which suggests that the efficiency of information transfer in the brain's neural network is the same across mammals.
Researchers at Graz University of Technology developed a new machine learning algorithm called e-prop, which significantly expands the possible applications of AI. This novel approach uses spikes to enable more efficient information processing and reduces energy consumption.
Zhao and Cheng are working on a project to develop new gradient-free methods for training various types of deep neural networks. They aim to create an algorithmic and theoretical framework for model parallelization based on gradient-free optimization, as well as efficient distributed workflow systems.
Researchers at Medical University of South Carolina found that the brain uses similar visual areas for mental imagery and vision, but with less precision. This study has potential applications for understanding PTSD and other mental health disorders affecting mental imagery.
Researchers analyzed hundreds of thousands of Instagram posts about vaping and found that 40% were promoting flavored e-liquids to young audiences. The study highlights the need for stricter laws and regulations on social media advertising targeting younger users.
Researchers from North Carolina State University discovered that incorporating Hamiltonian function into neural networks enables them to better predict and respond to chaos. This innovation has significant implications for improved artificial intelligence applications.
Researchers from Lobachevsky University and international colleagues have developed a model that demonstrates the existence of concept cells in the brain, which can process and learn abstract concepts. The study suggests that individual neurons, rather than large neuronal complexes, are responsible for complex tasks performed by humans.