Articles tagged with Artificial Neural Networks
Researchers at the University of Bonn have developed a new training technique for highly efficient AI methods, inspired by biological neurons that use short voltage pulses to communicate. This approach enables spiking neural networks to be trained using conventional methods, resulting in improved accuracy and reduced energy consumption.
DNNs have an inbuilt 'Occam's razor,' favouring simpler solutions that fit training data. This bias helps them generalize well on simple patterns but may struggle with complex data, aligning with real-world data characteristics.
The study reveals that directional connections propagate signals in a downstream flow, leading to more complex activity patterns. Mathematical models also suggest that modularity and connectivity interact to foster dynamical complexity.
A new method called Annotatability helps identify mismatches in cell annotations and better characterizes biological data structures. This approach enables more precise downstream analysis of biological signals, capturing cellular communities associated with target signals.
Current energy-hungry transformer-based systems contrast with Turing's idea of machines that develop intelligence naturally, like human children. AI systems can now perform tasks exclusive to human intellect, such as generating coherent text and discussing abstract ideas, but with limitations on sustainability and societal impact
Researchers at KAIST developed a new method to learn without weight transport, enabling faster and more accurate learning. By pre-training with random noise, the team showed that neural networks can achieve high learning efficiency and solve the weight transport problem.
Researchers developed a new benchmark for health care using reinforcement learning, which shows promise in managing chronic or psychiatric diseases. However, current methods are data-hungry and fail to perform accurately when tested on real-world data.
A groundbreaking AI model called NitroFusion creates images in seconds using modest hardware, eliminating the need for large computing resources. The open-source technology enables creative professionals and individuals to produce high-quality images affordably.
Physicists at the University of Michigan have developed an algorithm that enables materials to learn and adapt, mimicking brain-like behaviors. This breakthrough has implications for the development of advanced materials with self-tuning properties.
A novel classification method for adult spinal deformity diseases has been developed using deep learning of gait data, achieving a correct response rate of 71.43% in testing, surpassing conventional methods.
Scientists at MIT developed a fully integrated photonic processor that can perform all key computations of a deep neural network optically on the chip. The device completed machine-learning classification tasks in under half a nanosecond while achieving over 92% accuracy, similar to traditional hardware.
Researchers developed an AI tool called BrainBench to test large language models' ability to predict neuroscience study outcomes. The results showed that LLMs surpassed human experts with an average accuracy of 81%, highlighting their potential as powerful tools for accelerating research.
Recent Nobel Prizes in physics and chemistry have recognized the convergence of AI with physics and chemistry, emphasizing the need for interdisciplinary research. Researchers advocate for nurturing AI-enabled polymaths to bridge the gap between theoretical advancements and practical applications.
Researchers at Cold Spring Harbor Laboratory have devised a potential solution to the paradox of animal innate abilities using artificial intelligence. The genomic bottleneck algorithm allows for compression levels unseen in AI, enabling faster runtimes and potentially leading to more evolved AI systems.
A new AI tool generates realistic satellite images of future flooding, which can help communities visualize and prepare for approaching storms. The method combines a generative artificial intelligence model with a physics-based flood model, producing more accurate and realistic images than an AI-only approach.
Researchers have trained AI models to distinguish brain tumors from healthy tissue using convolutional neural networks and transfer learning. The models achieved an average accuracy of 85.99% at detecting brain cancer, with the ability to generate images showing specific areas in its tumor-positive or negative classification.
Researchers propose quasi-convolution coding to simplify reservoir computer design, enhancing memory capacity and reducing complexity. The approach leverages dual polarization modes of commercial lasers, offering a feasible strategy for constructing integrated deep RC systems.
Researchers suggest that starting with smaller neural networks and using curriculum learning can improve performance and reduce the need for massive computing resources. This approach could lead to more resource-efficient and less energy-consuming AI systems.
Researchers found that membership inference attacks on large language models (LLMs) are not effective in measuring information exposure risks. The common method used to test LLM leaks suffers from ambiguity due to the fluidity of language, making it difficult to define a representative set of non-member candidates.
A deep learning AI model can identify pathology in images of animal and human tissue much faster and often more accurately than people, potentially revolutionizing disease-related research and medical diagnosis. The model was trained using images from past epigenetic studies and showed accuracy comparable to human experts.
Researchers developed a novel AI approach to predict atomic-level chemical bonding information in 3D space, bypassing traditional supercomputer simulations. This methodology accelerates calculations by learning chemical bonding information using neural network algorithms from computer vision.
The AI-powered system can detect toxic gases like nitrogen dioxide in real-time, identifying the source of harmful gas leaks. The system's optimization technique ensures fewer resources are used while providing faster and more accurate gas leak detection.
Research on optical neural networks (ONNs) has made significant progress, addressing challenges of low integration, stability, and portability. ONNs offer advantages over modern computing hardware, enabling strong computational support for societal development.
A new training algorithm called ternarized gradient BNN (TGBNN) enables learning capabilities for binarized neural networks (BNNs) on IoT edge devices. The proposed MRAM-based CiM architecture achieves faster convergence and matching accuracy with regular BNNs.
A machine learning model predicts soil behavior during earthquakes, identifying areas vulnerable to liquefaction and providing contour maps for safer construction sites. The study uses geological data to create detailed 3D maps of soil layers, improving prediction accuracy by 20%.
Researchers have discovered a ferroelectric material that can adapt to light pulses on the nanoscale, creating networked nanodomains that can be reconfigured without requiring much energy. This discovery could lead to more energy-efficient computing systems and artificial neural networks.
Researchers at Chung-Ang University developed a novel GAN model, PMF-GAN, to address stability and efficiency issues. The model utilizes kernel functions and histogram transformations to improve the generator's ability to produce diverse outputs, reducing mode collapse and gradient vanishing.
Researchers developed an electronic tongue that can identify differences in liquids and detect food safety concerns. The AI-powered system achieved high accuracy when using its own assessment parameters, providing insights into the neural network's decision-making process.
Researchers propose a new approach to reduce the tradeoff between overhead and protecting machines against vulnerabilities. The 'Vulnerability-Adaptive Protection Paradigm' applies different protection strategies to different parts of the system, allocating resources more wisely.
Researchers at TU Graz have developed a new machine learning method that generates precise live MRI images of the beating heart using only a few MRI measurement data. This breakthrough enables faster and cheaper MRI applications, including quantitative MRI for diagnoses.
Researchers developed DIAMANTE, a data-centric semantic segmentation approach to detect forest tree dieback events in satellite images. The approach trains a U-Net-like model on labelled remote-sensing datasets and achieves reasonable accuracy for early disease detection, reducing false alarms.
Neuroscientists have discovered a global process across the brain that coordinates sensory input with motor action through learning. In trained mice, neurons link sensory evidence to action initiation, integrating information across multiple brain regions.
Using AI and the connectome, researchers can now predict individual neuron activity in living brains. The new model predicts neural activity in response to visual input and accurately reproduces over two dozen experimental studies.
Researchers developed a computational method called DISCOVER to break down images into semantically meaningful components that AI uses to make decisions. The technology demonstrates the interpretation of AI decisions for various medical imaging tasks, including IVF embryo analysis and Alzheimer's brain imaging.
Researchers developed an AI model to detect lung disease in premature babies by analyzing their breathing patterns while sleeping. The Long Short-Term Memory (LSTM) model achieved 96% accuracy in classifying flow values as belonging to a patient with BPD or not, enabling early diagnosis and treatment.
A deep-learning algorithm developed by astronomer David Harvey can untangle the complex signals of self-interacting dark matter and AGN feedback in galaxy cluster images. The Inception model achieved an accuracy of 80% under ideal conditions, showcasing its potential for analyzing vast amounts of space data.
Researchers created a data set of over 10 million documents to test detection ability in current and future detectors. They found that most detectors only work well in specific use cases and can be easily evaded by manipulating the text. The new tool, RAID, aims to provide a standardized benchmark for robust detection.
Researchers developed a framework called SigLLM that uses large language models to detect anomalies in time-series data. The approach converts time-series data into text-based inputs and can be deployed right out of the box, offering an efficient anomaly detection solution for complex systems.
EPFL researchers have created an energy-efficient method for nonlinear computations using scattered light from low-power lasers. The new approach is scalable and up to 1,000 times more power-efficient than state-of-the-art digital networks, making it suitable for realizing optical neural networks.
A new research consortium aims to improve the reliability of machine learning systems by using geometric methods to prevent adversarial attacks. The project, GeoMAR, will explore ways to feed neural networks with erroneous data during training to prepare them for real-world scenarios.
Researchers at Pohang University of Science & Technology have developed a novel analog hardware using ECRAM devices that maximizes AI computational performance. Their technique, which uses a three-terminal structure with separate paths for reading and writing data, demonstrates excellent electrical and switching characteristics.
A new study published in the journal Brain Connectivity reveals how psychological resilience can aid children's recovery from concussions. The research found that building resilience through supportive family environments and effective coping strategies may help young patients heal faster.
Researchers at Max Planck Institute propose a new method for implementing neural networks with optical systems, which could lead to faster and more energy-efficient alternatives. The approach allows for parallel computations in high speeds limited by the speed of light, and can be applied to various physically different systems.
UCF's STRONG-AI initiative aims to uplift bright, low-income undergraduate students in pursuing well-rounded AI education through faculty and peer mentorship and scholarship. The program has received over 150 applications and will select 10-15 students annually based on financial aid eligibility and academic success.
A research team at Pusan National University proposes a novel backscatter communication system that utilizes transfer learning and polarization diversity to achieve 40% energy efficiency gains compared to conventional systems. The system enables integrated sensing and communication technology, facilitating smart cities, efficient indus...
Researchers create an analog system that can learn complex tasks like XOR relationships and nonlinear regression, using local learning rules without centralized processor. The system is fast, low-power, and scalable, offering a unique opportunity for studying emergent learning.
A new machine learning technique called Dual-Channel Prototype Network (DCPN) can efficiently classify pathological images with limited data, which is essential for diagnosing rare diseases. The DCPN uses few-shot learning to make predictions and achieves noticeable advantages over other methods on three public datasets.
A research team at KAIST has developed an AI-based methodology to predict the major elemental composition and charge-discharge state of NCM cathode materials with high accuracy using convolutional neural networks. The technology can analyze surface morphology images of batteries to determine their composition and lifespan.
A team of Chinese researchers introduced CNNs into optics, developing an ultrafast convolutional optical neural network (ONN) for efficient and clear imaging. The ONN achieves true optical computing speeds, significantly enhancing image quality and enabling real-time dynamic imaging.
Scientists have developed an AI that can navigate new environments, seek rewards, map landmarks and overcome obstacles using a novel approach inspired by the brain circuits of sea slugs and octopuses. The new AI, called CyberOctopus, has the ability to explore and gather information while learning on the job.
Researchers investigated the efficiency of modern neural network-based generative models, comparing them to traditional sampling techniques. The study found that modern diffusion-based methods may face challenges due to a first-order phase transition, but also exhibit superior efficiency in certain cases.
A new model developed by Flatiron Institute researchers proposes that individual neurons exert more control over their surroundings, which could be replicated in artificial neural networks. This updated model treats neurons as tiny 'controllers' and may lead to better AI performance and efficiency.
A new study published in GEN Biotechnology describes the establishment of a 3D hydrogel-based platform for producing functional T-cells from hematopoietic stem and progenitor cells. The platform was engineered with key thymic components to direct T-cell development, producing cytokine-producing T-cells.
Researchers at Bar-Ilan University have discovered a new scaling law that governs how artificial neural networks handle an increasing number of categories for identification. This law reveals how the identification error rate increases with the number of required recognizable objects, impacting AI latency and efficiency.
Researchers created a virtual rodent with an artificial brain that can move like a real rodent, simulating complex behaviors using biomechanical data from real rats. The model uses deep reinforcement learning and AI to predict neural activity across behaviors.
A new study reveals that ChatGPT's automated content moderation filters can flag nearly 20% of its own generated scripts for content violations, including half of PG-rated shows. The research raises questions about the efficacy of using AI as a tool in scriptwriting and its potential impact on artistic expression.
A new study from the University of Tsukuba introduces an algorithm that determines the application ratio of various compression methods for minimizing data amount in CNNs. This leads to a 28 times smaller model and 76 times faster computation compared to previous models.
Researchers found human infants use 'helpless' period to pre-train brain, leading to rapid learning and high performance, similar to machine learning models. This study challenges classic explanation for infant helplessness and could inspire next gen AI models.
Researchers developed an innovative scheduling system for electric vehicles that enhances power grid efficiency by synchronizing charging with peak solar energy production times. The system reduces energy loss, prevents power outages, and minimizes the impact of EV charging on the grid.
A study published in PLOS Biology found that negating adjectives affects brain processing, with slower reaction times and varied interpretations. The researchers used magnetoencephalography to capture brain activity while participants rated affirmative or negated phrases.