Articles tagged with Neural Networks
A new method accelerates template creation for medical-image analysis, generating brain scan templates based on patient attributes such as age and sex. The model can synthesize atlases from sparse data, improving disease diagnosis accuracy.
Researchers from Russia and Germany discovered that activation of GDNF helps protect brain cells from death during hypoxic damage, maintaining neural network activity. This finding can lead to the development of an effective method for correcting CNS pathologies developing under oxygen deficiency.
Researchers at Argonne National Laboratory have developed domain-aware neural networks to replace expensive parameterizations in the Weather Research and Forecasting (WRF) model. These algorithms can predict environmental data more accurately with significantly less training data, enabling faster and higher-resolution simulations.
Researchers used deep neural networks to analyze ECG test results from over 2 million patients, identifying those at high risk of developing atrial fibrillation or dying within a year. The models were found to be superior in predicting mortality risk even in patients with normal ECGs.
Researchers developed a novel computational approach using deep artificial neural networks to predict neural responses to images. The study found that certain stimuli, such as checkerboards or sharp corners, elicit strong responses from neurons, contradicting current dogma in the field.
Researchers at Duke University have trained an AI tool to identify up to 200 species of birds from just a photo. The system, which uses deep learning, also shows its thinking by highlighting key patterns in the image.
Researchers developed an AI-based ozone forecasting system that can predict ozone levels with 85-90% accuracy. The model uses convolutional neural networks to analyze current conditions and forecast future ozone levels, improving health alerts for people at risk.
A new theory suggests that consciousness arises from synchronized neural activity and is guided by thermodynamic principles. During conscious states, the brain has higher entropy and more connected neural networks, leading to greater mental flexibility.
A team of researchers has successfully mapped the local connectome in the cerebral cortex using 3D electron microscopy, producing a connectome about 26 times larger than previous ones. The study provides insights into the density and magnitude of neuronal networks in the brain.
Researchers have developed an artificial intelligence technique that uses deep neural networks to analyze data from experiments on nanoscale ferroelectrics. This method has identified geometrically-driven differences in ferroelectric domain switching, providing new insights into the mechanisms of ferroelectric switching.
Researchers developed a neural network, PatchFCN, trained on 4,396 CT scans to detect brain hemorrhage abnormalities with accuracy similar to human experts. The algorithm achieved high accuracy and pixel-level delineation, classifying abnormalities into different pathological subtypes.
Researchers developed lipid-based memcapacitors that mimic biological synapses, accelerating routes to neuromorphic computing. The discovery could support the emergence of biology-inspired computing networks for sensory approaches to machine learning.
Researchers at the University of Delaware are developing new memory devices that can support neural networks in low-power embedded systems. These advancements aim to improve the lifetime and reliability of IoT devices, which currently struggle with battery power and memory constraints.
Researchers used repetitive transcranial magnetic stimulation to increase functional connectivity of a neural network implicated in memory. The study, published in eNeuro, confirms the effectiveness of this technique for experimental and clinical applications.
Researchers at Duke University use machine learning to model complex biological circuits, achieving speeds of hours instead of years or months. By training a deep neural network on large datasets, they uncover patterns and interactions between variables that were previously impossible to discover.
RUDN University mathematicians developed a model to optimize data center efficiency using Markov chains. Their method reduces server overheating and improves server capacity utilization, resulting in significant cost savings.
Duke University engineers used machine learning to design dielectric metamaterials that absorb and emit specific frequencies of terahertz radiation, reducing calculation time from over 2,000 years to just 23 hours. The new designs enable thermophotovoltaic devices that convert waste heat to electricity with higher efficiency.
Researchers have developed an electronic chip that can perform high-sensitivity intracellular recording from thousands of connected neurons simultaneously. This breakthrough has enabled the mapping of hundreds of synaptic connections and opens up new strategies for machine intelligence to build artificial neural networks.
A highly predictive genetic risk score is being developed by Paul Tran to identify children at significant risk of developing type 1 diabetes. The algorithm uses a feedforward neural network to analyze thousands of gene variants associated with the disease, aiming to predict with five times better accuracy than current systems.
Researchers have developed a brain-inspired, analog neural network that provides probabilistic responses for complex decision-making. The device is more energy efficient and produces less heat than current computing architectures.
Researchers have developed a new gene therapy that converts glial cells into neurons, improving motor function in mice and potentially treating stroke. The treatment uses the NeuroD1 gene and has been shown to increase neuronal density and reduce brain tissue loss in mouse models of stroke.
Researchers at Max Planck Florida Institute for Neuroscience developed a strategy to label and map local inhibitory inputs onto cells. They found that inhibitory inputs may parallel or diverge from target neurons, revealing a diverse palette of inhibition. This discovery suggests complex functional connectivity in the visual cortex.
A study reveals that alternative splicing controls the identity and function of nerve cells, allowing for a complex neuronal network with limited genes. The research team mapped splice variants in different types of neurons, identifying unique repertoires that shape their characteristics.
Researchers use TDA to inject knowledge of real world into neural networks, reducing training time and increasing intelligibility. This approach enables machines to focus on meaningful features and improve performance in tasks like face recognition.
A two-layer all-optical artificial neural network has been successfully demonstrated for complex classification tasks, outperforming computer-based neural networks. The researchers plan to expand this approach to large-scale optical deep neural networks for specific practical applications.
Scientists have successfully grown miniature brains from stem cells that exhibit functional neural networks and produce brain waves resembling those of preterm babies. The study marks a significant breakthrough in understanding human brain development and may lead to new insights into diseases such as autism, epilepsy, and schizophrenia.
Researchers from MIPT created a second-order memristor that stores information and forgets it over time, mimicking natural memory. The device is based on hafnium oxide and has potential applications in designing analog neurocomputers.
The UCLA researchers have significantly increased the system's accuracy by adding a second set of detectors to the system, representing each object type with two detectors rather than one. The new design takes advantage of parallelization and scalability of optical-based computational systems.
Researchers have developed an all-optical diffractive neural network that achieves unprecedented levels of inference accuracy, closing the performance gap with electronic neural networks. The design incorporates a differential detection scheme, which enables specialized sub-networks to recognize specific object classes.
Researchers at Purdue University have developed a new software called Emap2sec that can identify secondary structures in proteins from lower-resolution cryo-EM maps. This technique has the potential to speed up protein structure analysis and improve accuracy, enabling researchers to develop more effective drugs for various diseases.
A new computer system called EmoNet can accurately categorize images into emotional categories, suggesting that the visual cortex plays a crucial role in emotion processing. The study found that EmoNet could recognize emotions with high accuracy, even for nuanced emotions like confusion and awe.
Researchers from Forschungszentrum Jülich and RWTH Aachen University have identified a second critical mode in neuronal networks, allowing for parallel information processing. This newly discovered dynamics permits the network to represent signals in numerous combinations of activated neurons.
A new deep neural network architecture can identify manipulated images at the pixel level, detecting blurred boundaries and unnatural transitions between regions. This technology aims to improve photo editing tool security and detect deepfakes with high precision.
Scientists from Russia and Greece have successfully implemented a spiking neural network based on memristors, demonstrating the feasibility of local learning rules. The research enables autonomous unsupervised learning of complex neural networks, paving the way for new applications in AI.
Researchers from the University of Barcelona successfully synchronized two nanoscale optomechanical oscillators through mechanical coupling. The study demonstrates collective dynamics that can be controlled by acting externally on one oscillator only.
Researchers developed a new computational method using neural networks to simulate open quantum systems, predicting properties of large-scale quantum systems. This approach addresses the challenges of simulating intrinsically complex tasks with exponentially growing computational power.
Scientists have created functional neural networks derived from cerebral organoids, which can mimic the development of the human brain. The study provides a new tool for understanding brain function and may lead to breakthroughs in drug discovery, modeling neuropsychiatric disorders, and regenerative medicine.
Researchers used AI to train neural networks on complex behavioral tasks, revealing two distinct processes involved in short-term memory. These processes include a 'silent' process where the brain stores memories without ongoing neural activity, and a more active process where circuits of neurons fire continuously.
A recent study from MIT has found that dendrites are nearly always active when the main cell body of a neuron is active, suggesting a larger role in neural computation. The researchers used calcium imaging to measure activity in both soma and dendrites of individual neurons in the visual cortex.
A team of computer scientists has developed a method to analyze artistic portraiture, capturing facial features and individual style with high accuracy. The researchers used 'artistic augmentation' to transform photographic face data into more similar to artistic portraits.
Researchers developed a low-cost, sensor-packed glove that enables an AI system to recognize objects through touch alone. The glove produced high-resolution data at a fraction of the cost of existing sensors, allowing for accurate object classification and weight prediction with up to 76% accuracy.
Researchers studied inhibitory neurons' impact on brain oscillations using computer models. They found that these neurons can delay or facilitate the onset of synchronization, which is crucial for understanding brain diseases like Alzheimer's and epilepsy.
A study published in PNAS finds that premature infants who listened to tailored music had improved brain network development and functional connectivity compared to those without music. The research suggests that music can be a valuable tool to support the development of fragile newborns.
Research suggests both single neurons and large neural assemblies support spatial navigation; a potential link has been discovered between the two, with EEG oscillations potentially constituting the connection.
Researchers used zebrafish to study the effects of a genetic mutation linked to brain disorders such as autism and schizophrenia. The study found that the mutation caused clustering of cellular interactions, leading to disrupted normal development and brain health.
Researchers at North Carolina State University have developed AOGNets, a new framework for building deep neural networks via grammar-guided network generators. The new networks outperformed existing state-of-the-art frameworks in visual recognition tasks, achieving better prediction accuracy and model interpretability.
Researchers have developed a machine learning model to rapidly predict plasma behavior, allowing for real-time control of fusion reactions on Earth. The new model reduces calculation time from minutes to microseconds, enabling faster decision-making during experiments.
A team of researchers developed CosmoGAN, a deep learning network that generates high-fidelity convergence maps for weak gravitational lensing. The model achieves high statistical agreement with fully simulated maps, paving the way for building emulators out of deep neural networks.
Researchers studied embryonic development to understand how neurons regulate digestive movement. They discovered that intestinal nervous system coordinates muscular contractions and reflexes.
Researchers have developed a new framework that enables deep neural networks to learn new tasks while minimizing the loss of previously learned information. The Learn to Grow framework demonstrates improved performance in both new and old tasks, with backward transfer occurring when learning a new task enhances previous task accuracy.
A team of researchers from the Universities of Münster, Oxford, and Exeter have developed a light-based hardware that mimics the behavior of neurons and synapses in the brain. The chip can process data much faster than traditional computers, enabling applications such as medical diagnoses and cancer cell identification.
Scientists have developed a new method to quickly map brain connections, enabling systematic study of connection patterns within single individuals. The technique, combining infrared laser stimulation with functional MRI, reveals the direction of information flowing in the brain, critical for understanding brain processing.
Studies found that all brain parts process touch signals, complementing each other for perception. Brain network processing information as a single network with partially different functions from situation to situation.
Researchers have found that melanin-concentrating hormone neurons are active during rapid-eye movement (REM) sleep and when exploring novel objects in mice. This suggests these cells may facilitate memory formation through single-cell activity patterns.
A team of scientists at MIT developed a neural network that can read scientific papers and generate a plain-English summary. The system, called RUM, uses vectors rotating in multidimensional space to represent words and improve memory and recall capabilities.
A new neural network system developed by Stanford researchers enables autonomous cars to learn from past driving experiences and adapt to unknown conditions. The system performed similarly well as an experienced racecar driver in high-friction and low-friction scenarios, showing promise for improved safety.
Researchers developed a new 'multi-z' confocal microscopy system for imaging large groups of cells, enabling fast and detailed imaging across a wide field of view. The instrument captured cellular details at high speeds over a large 3D volume, providing unprecedented insights into how neurons interact during various behaviors.
Researchers at MIT have developed a neural architecture search algorithm that can directly learn specialized convolutional neural networks for target hardware platforms in only 200 GPU hours. The algorithm uses 'path-level' binarization and pruning to reduce memory consumption and improve efficiency.
Researchers at Harvard University have developed a novel brain implant that mimics the appearance, size, and flexibility of real neurons, allowing for stable monitoring of neural signals and potential treatment of neurological disorders. The implants inspire negligible immune response and may even encourage tissue regeneration.
Researchers developed a synchronization registration method with high sensitivity and selectivity, enabling the network to recognize up to 14 figures out of 102 possible variants. The system operates independently as a separate neural organism, utilizing multilevel neurons with high functionality.