Articles tagged with Neural Networks
A new theory of dreaming proposes that our brains create strange dreamscapes to counteract overfitting by familiarizing themselves with everyday experiences. This 'overfitted brain hypothesis,' inspired by AI regularization techniques, suggests dreams serve a purpose in generalization and world understanding.
Politecnico di Milano researchers used neural networks to predict the acoustic behavior of violin plates based on geometric parameters. The results showed an accuracy close to 98%, enabling luthiers to design and build violins with optimal sound quality, exploring new designs and materials.
A scientist from HSE University has developed an image recognition algorithm that speeds up real-time processing of video-based image recognition systems by up to 40%. The algorithm uses a sequence of convolutional layers and fine-tuning to achieve accurate results while controlling loss in accuracy.
A study found that deep neural networks can accurately predict lung cancer type from CT scans, identifying new associations between genes and imaging features. This approach increases radiologists' confidence in assessing tumor types, informing individualized treatment planning.
Researchers at Carnegie Mellon University have developed a technique using machine learning and high-performance computing to simulate complex universes in less than a day. The approach enables high-resolution cosmology simulations, advancing physics research and providing new insights into the universe's mysteries.
Researchers used neural networks to simulate complex universes, reducing computation time by a thousandth. The new method allows for both high resolution and large volume simulations, holding the potential for major advances in numerical cosmology and astrophysics.
A novel auditory perception model simulates human ear dynamics to capture time dynamics of dimensional emotions. Neural networks then extract features that reflect this time dynamics, showing better emotion recognition performance than traditional acoustic-based features.
General Motors has licensed the award-winning AI software system MENNDL from Oak Ridge National Laboratory to accelerate advanced driver assistance systems technology and design. MENNDL uses evolution to design optimal convolutional neural networks, dramatically speeding up the process of recognizing patterns in datasets.
A research team created a computer model that can simulate the impact of individual receptor types on brain activity. The model uses data from three imaging techniques to quantify receptor-specific modulations of brain states. By predicting changes in brain dynamics after receptor activation, the researchers hope to develop new diagnos...
A new real-time 3D motion tracking system combines transparent light detectors with advanced neural network methods to enable fast tracking speed, compact hardware, and lower cost compared to existing solutions. The technology has promising applications in automated manufacturing, biomedical imaging, and autonomous driving.
Researchers developed DeepShake, a deep spatiotemporal neural network trained on over 36,000 earthquakes, which analyzes seismic signals in real time and provides advanced warnings of strong shaking. The model was tested using the 2019 Ridgecrest earthquake, sending simulated alerts up to 13 seconds prior to high-intensity ground shaking.
A team of researchers has developed an AI agent called Crystallography Companion Agent (XCA) to analyze X-ray diffraction data and identify material properties faster. The agent collaborates with scientists to perform autonomous phase identifications, overcoming traditional neuronal network overconfidence.
Researchers at KAUST developed a brain-on-a-chip that can learn real-world data patterns without extensive training, leveraging spiking neural networks and spike-timing-dependent plasticity model. The system is more than 20 times faster and 200 times more energy efficient than other neural network platforms.
A neural network developed by Skoltech researchers improves credit scoring using transactional banking data, surpassing existing models. The EWS-GCN model processes large-scale temporal graphs directly and aggregates information to predict target client credit ratings.
Researchers created a satellite-based map of human pressure on lands around the world using machine learning. The map reveals abrupt changes in landscapes due to deforestation, mining, and urbanization, providing insights into biodiversity conservation and sustainability.
Researchers found altered brain activity in individuals with chronic sinusitis, affecting neural networks that modulate cognition and response to external stimuli. Despite no significant clinical impairment, participants showed subtle brain region communication changes associated with attention decline and sleep disturbances.
Researchers found that neural networks trained on sound files of human language reached higher performance in image recognition, identifying objects and animals correctly 92% of the time. Using sound as a training tool improved results even with limited training data, outperforming traditional binary input methods.
A new deep neural network architecture can differentiate between healthy and diseased skin images with high accuracy, offering a potential screening tool for systemic sclerosis. The proposed network reached 100% accuracy in training and validation sets, outperforming traditional CNNs.
Researchers have created an early prototype of a medical imaging system using neural networks to analyze near-infrared images of veins and project a venous pattern onto a patient's body. The system can detect vein contours accurately, fully automatically, and independently, reducing discomfort for patients with difficult access to veins.
A new virtual diagnostic approach uses machine learning to analyze beam quality in electron microscopes, X-ray lasers, and medical accelerators. The method provides accurate information that conventional diagnostics cannot, enabling operators to optimize device performance.
A new recurrent neural network framework enables fast and efficient 3D imaging of fluorescent samples, reducing scan times by ~30-fold. The approach uses few 2D images to reconstruct 3D images, mitigating photo-bleaching challenges in live sample experiments.
Artificial neurons help decode cortical signals using a new algorithm that automates feature extraction and interpretation. The neural network architecture is automatically tuned to analyze signals from separate neural populations, providing physiologically meaningful results.
MIT researchers develop a deep-learning algorithm to optimize sensor placement on soft robots, allowing them to better interact with their environment and complete assigned tasks. The algorithm learns the most efficient sequence of movements and identifies the most important particles to improve performance.
The study reveals how neural circuits balance excitation and inhibition, crucial for normal functionality of our brain. The results provide a clearer picture of how this balance is preserved and where it fails in living neural networks.
A novel convolutional neural network, FMNet, was developed to determine the source focal mechanism of earthquakes rapidly using full waveforms. The method proves effective in calculating parameters within one second with minimal computing resources.
Researchers at UC San Diego have developed a nanoscale artificial neuron device that efficiently carries out activation functions in hardware, reducing computing power and circuitry. The device, which implements the rectified linear unit activation function, can process images and perform edge detection with high accuracy.
A new algorithm, C2FIV, uses facial motion to verify identities, providing an additional layer of security. With a success rate of over 90% accuracy in its preliminary study, the technology has broader applications beyond smartphone access, including workplace and online banking security.
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 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 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 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.
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.
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.