Articles tagged with Artificial Neural Networks
MIT researchers develop a method to test feature-attribution methods for machine-learning models. They find that even the most popular methods often miss important features in an image and some perform as poorly as a random baseline. This has major implications for high-stakes situations like medical diagnoses.
Researchers at the University of Surrey have successfully demonstrated the use of multimodal transistors in artificial neural networks, achieving practically identical classification accuracy as pure ReLU implementations. The study paves the way for thin-film decision and classification circuits, which could be used in more complex AI ...
Researchers at RIKEN CBS demonstrate that neural networks minimize energy cost and solve mazes efficiently, pointing to a set of universal mathematical rules. The findings will aid in analyzing impaired brain function and generating optimized neural networks for artificial intelligences.
A team of researchers from Osaka University has developed a simple system based on electrochemical reactions that can perform complex calculations. The system uses polyoxometalate molecules and deionized water to process information and solve nonlinear problems.
A research team at SUTD has developed an ultra-scalable artificial synapse using 2D materials, enabling the commercialization of brain-inspired hardware. The device integrates functional and silent synapses into a single unit, reducing hardware costs and improving efficiency.
Researchers at MIT and Google Brain developed a system that predicts how changing materials or designs will improve solar cell performance. The new simulator, called differentiable solar cell simulator, provides information on which changes will provide desired improvements, increasing the rate of discovery of new configurations.
A team of scientists has created a neural network that can predict and generate new protein structures using deep learning. The network, trained on random protein sequences, can produce stable protein shapes with remarkable accuracy.
Scientists at TU Wien have developed a novel germanium-based transistor with the ability to perform different logical tasks, offering improved adaptability and flexibility in chip design. This technology has potential applications in artificial intelligence, neural networks, and logic circuits that work with more than just 0 and 1.
A team of researchers, including those from Rensselaer Polytechnic Institute and the University of Washington, have developed a neural network that can predict protein shapes with high accuracy. The network was trained on random protein sequences and generated 2,000 new proteins, many of which were successfully produced in the lab.
A new confocal platform using artificial intelligence and multiple lenses improves volumetric resolution by over 10-fold while reducing phototoxicity. The platform uses Deep Learning algorithms to distinguish between high-quality images with low signal-to-noise ratio and better images.
Researchers discovered the retrosplenial cortex as the site of value decision-making in the brain. Persistency allows value signals to be effectively represented across different brain areas, especially the RSC. Artificial intelligence networks mimicking mouse decisions showed remarkably similar results.
A new study by USC researchers uses GANs to generate synthetic neurological data that can be fed into machine-learning algorithms to improve BCI usability. This approach improved BCI training speed by up to 20 times and enabled rapid adaptation to new subjects.
Researchers trained an artificial intelligence algorithm to predict the next designer drugs before they are even on the market, allowing law enforcement agencies to identify and regulate new versions of dangerous psychoactive drugs. The model was tested against 196 new designer drugs and found nearly all were present in its generated set.
A new machine learning-based approach enhances student engagement in online environments. The algorithm detects when students disengage, prompting interventions to improve learning outcomes.
A new study explores the problem of shortcuts in a popular machine learning method and proposes a solution that can prevent shortcuts by forcing the model to use more data. By removing simpler characteristics and asking the model to solve the task two ways, researchers reduce the tendency for shortcut solutions and boost performance.
A new algorithm has been developed to train spiking neural networks, mimicking the human brain's structure and function. This approach enables these powerful, fast, and energy-efficient systems to solve complex tasks like image classification with high precision.
Researchers have trained a neural network to detect anomalies in medical images, adapting it to the nature of medical imaging and achieving better results. The new method uses weakly supervised training and can spot small-scale anomalies, accelerating the work of histopathologists and radiologists.
Researchers at Osaka University developed a deep neural network to accurately determine qubit states despite environmental noise. The novel approach may lead to more robust and practical quantum computing systems.
Researchers developed an attention-based deep neural network to detect multiple ship targets, exceeding conventional networks' performance. The model focused on inherent features of the two ships simultaneously, outperforming traditional approaches.
A team of scientists at NAIST successfully used automatic differentiation to accelerate calculations of model parameter extraction, reducing computation time by 3.5 times compared to conventional methods. This breakthrough enables the design of more efficient power converters with increased performance and reduced energy consumption.
A team of researchers from the University of Illinois Urbana-Champaign used advanced machine learning to model the physico-chemical properties of a molten salt compound called FLiNaK, enabling accurate atomic-scale reproduction and prediction of behavior under specific reactor conditions. This computational framework can help character...
Researchers at Singapore University of Technology and Design (SUTD) have designed an ultralow power artificial synapse for next-generation AI systems. The team's innovation uses a nanoscale deposit-only-metal-electrode fabrication process, achieving an all-time-low energy consumption of 1.8 pJ per pair-pulse-based synaptic event.
Scientists discovered that recurrent neural networks (RNNs) play a crucial role in the frontal cortex, responsible for decision-making, expressive language, and voluntary movement. The research also found that RNNs are more complex than previously thought, with a unidirectional structure.
Researchers have developed a next-generation reservoir computing that solves complex problems in less than a second, compared to current supercomputers. The new system uses significantly fewer computing resources and less data input, making it 1 million times faster for accurate forecasts.
Scientists developed a machine learning method to analyze NMR data, allowing faster and more accurate analysis of proteins and chemical reactions in the human body. The method uses an artificial deep neural network to separate and analyze complex data, resulting in highly reproducible results comparable to human experts.
Researchers at Osaka University used machine learning to analyze locomotion data from diverse species, revealing common features associated with dopamine deficiency. The study found that worms, mice, and humans exhibit similar movement disorders when lacking dopamine, despite their evolutionary differences.
A study by Purdue University and collaborators has found a way to demonstrate habituation and sensitization in nickel oxide, a quantum material that mimics the sea slug's most essential intelligence features. This discovery could lead to building hardware-based AI with improved efficiency and reduced energy consumption.
Researchers at The Hebrew University of Jerusalem have developed a new deep learning artificial infrastructure inspired by individual neurons. Their approach uses complex mathematical modeling to replicate the brain's electrical processes and create more intelligent AI systems.
A team of researchers at Osaka University created a custom dataset to train an AI algorithm to digitally remove unwanted objects from building façade images. The algorithm achieved high accuracy in inpainting occluded regions with digital inpainting.
Prof. Jae Youn Hwang's team developed an AI neural network module that can accurately extract buildings from aerial images for remote sensing. This technology can significantly improve the performance of extracting buildings from various aerial image domains.
Researchers at Technical University of Munich have developed a new machine learning algorithm that can analyze complex markets and their equilibrium strategies. This breakthrough has potential applications in auction theory, wireless spectrum auctions, and more.
The team used machine learning technique generative adversarial networks to digitally remove clouds from aerial images, generating accurate datasets of building image masks. This work may help automate computer vision jobs critical to civil engineering, enabling the detection of buildings in areas without labeled training data.
Researchers have created an AI software that uses Minecraft to test its ability to plan for future events and solve complex tasks. The software, developed by Penn State researchers, aims to advance artificial intelligence in areas such as robotics, logistics management, and drone flight.
A new study demonstrates that artificial intelligence networks based on human brain connectivity can perform cognitive tasks efficiently. Researchers created a brain connectivity pattern and applied it to an artificial neural network, which performed cognitive memory tasks more flexibly and efficiently than other benchmark architectures.
Researchers have created an artificial neuron that uses ions instead of electrons for information transmission, achieving a similar energy efficiency as the human brain. The device's ion channels and clusters replicate those found in neurons, allowing for the emission of action potentials and transmission of information.
Researchers at C-Crete Technologies have developed a method that utilizes deep learning to quickly predict and design novel hybrid organic-inorganic materials, offering improved materials design for various industries. By feeding quantum mechanics calculations to layered machine learning based on artificial neural networks, they can un...
A new study from Washington University in St. Louis shows that guided by sparsity, silicon neurons learn to pick the most energy-efficient perturbations and wave patterns, enabling an emergent phenomenon of efficient communication between neurons. This research has significant implications for designing neuromorphic AI systems.
A team at University of Notre Dame is using AI to transcribe ancient texts with high accuracy, improving capabilities of deep learning transcription. This project has significant implications for the digital humanities and historical archival research.
Researchers from the University of Groningen and Spain developed a method to train AI systems using distractions to improve image recognition. By analyzing how deep learning systems process images, they found that forcing the system's focus towards secondary characteristics can lead to better performance.
Neuroscientists used artificial intelligence to disentangle the relationship between perception and memory in the human brain. A novel computational framework predicts neural responses in the primate visual system, resolving decades-long debates over the role of the medial temporal lobe (MTL) in perception.
Researchers at Texas A&M University have developed a method to cool steam turbines using phase change materials, potentially reducing fresh water usage. By leveraging machine learning techniques, they created a system that can predict when and how much of the PCM will melt and freeze, maximizing cooling power and capacity.
Researchers from Skoltech and their colleagues developed a neural network that can efficiently generate IUPAC names for organic compounds in accordance with the IUPAC nomenclature system. The network, trained using the Transformer architecture, achieved an accuracy of nearly 99%, outperforming traditional rule-based solutions.
A researcher at MUSC has developed an AI algorithm that analyzes clinical notes to identify patients at risk of suicide. The algorithm achieved accuracy rates of around 98.5% when trained on electronic health records, and nearly 80% when validated against existing predictive models.
Scientists developed a machine learning algorithm that uses artificial neural networks to accurately forecast cell size as it grows and divides. By recognizing patterns in the data, the computer can make more complex predictions than conventional methods, which rely on simplifying assumptions.
The 'SynRap' project aims to accelerate the production of large amounts of synthetic data by a factor of one thousand using machine learning algorithms. The project will assess the quality of generated data sets in high energy density physics and high energy physics research areas.
Researchers from the University of Liège have developed a Bistable Recurrent Cell (BRC) that enables recurrent networks to learn temporal relationships over 1000 time steps, surpassing classical methods' limitations. This breakthrough could improve AI's ability to process time-series data and predict future events.
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.
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 recent study found that artificial neural networks develop spontaneous systems to name colours, adopting similar behaviours to humans. The researchers identified a universal property of optimizing complexity/accuracy trade-offs in discrete communication systems.
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 team of RIT researchers is working on developing an artificial intelligence system that can learn over time and play the popular video game Starcraft II. This project has the potential to advance practical solutions such as self-driving cars, service robots, and other real-world applications.
A team of researchers from UNIST developed a new learning method for PCM-based memristor neural networks, improving their learning ability by about 3% in handwriting classification tasks. The approach leverages the 'resistance drift' property of phase-change memory to update synapses and associate patterns with data.
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.
Researchers developed an AI model that uses ECG data to predict atrial fibrillation and AF-related stroke risk. The model identified high-risk patients with 62% accuracy, allowing for earlier intervention.
Researchers found similar properties between deep neural networks and primate visual cortices, providing insights into attention mechanisms. The study may accelerate AI development by understanding the neural basis of attention.
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.
A team of researchers has developed an optical computing core prototype using phase-change material, accelerating neural networks and reducing energy consumption for AI applications. The technology is scalable and directly applicable to cloud computing, making it a promising solution for the growing demands of AI online.
A breakthrough in optical neural networks accelerates computing speed and processing power to over 1000 times that of previous processors. The system can process record-sized images and achieve full facial image recognition.
Artificial neural networks enhance signal-to-background ratio in near-infrared imaging, sharpening blurred images. The technology has potential to improve diagnostics and image-guided surgery in the clinic.