Articles tagged with Artificial Neural Networks
Researchers have developed a system combining bio-inspired cameras with AI to quickly detect obstacles around cars, using less computational power. The hybrid system detects objects up to one hundred times faster than current systems while reducing data transmission and processing needs.
A team of neuroscientists developed a new AI algorithm that replicates the brain's visual processing, forming spatial maps and predicting sensory responses. The topographic deep artificial neural network (TDANN) could lead to more efficient artificial systems inspired by the brain's elegance.
Researchers at the University of Liège created a new type of spiking neuron, the Spiking Recurrent Cell (SRC), which combines simplicity with the ability to reproduce biological neuron dynamics. This innovation offers exciting prospects for neuro-inspired artificial intelligence, particularly in energy-efficient applications.
Researchers at the University of Michigan have created a new type of memristor that can mimic the timekeeping mechanism found in biological neural networks. This breakthrough could lead to significant energy savings for AI chips, potentially reducing energy consumption by a factor of 90 compared to current graphical processing units.
Model disgorgement is a set of techniques that force generative models to remove content leading to copyright infringement or biased responses. Researchers propose this approach to address issues like stylistic infringement, where models reproduce copyrighted works in the style of famous artists.
A new approach uses artificial intelligence to turn low-quality images into high-quality ones, enhancing the image quality of metalens cameras. This technology could make these cameras viable for intricate microscopy applications and mobile devices.
Researchers at Delft University of Technology developed a drone that flies autonomously using neuromorphic image processing and control based on the workings of animal brains. The drone's deep neural network processes data up to 64 times faster and consumes three times less energy than when running on a GPU.
Researchers created GraSSRep and rhea, tools that outperform current methods for handling repeats and structural variants in metagenomic data. These methods use self-supervised learning and graph neural networks to analyze microbiome data, offering new insights into biological processes and potential applications in antibiotic resistance.
Researchers developed a structure called multiplexed neuron sets to reduce crosstalk in optical neural networks. The new backpropagation training algorithm achieved comparable performance while improving energy efficiency by a factor of 10.
A new AI method developed by Swedish researchers can identify toxic substances based on their chemical structure, potentially replacing animal testing. The method has been shown to be more accurate and broadly applicable than existing computational tools, offering a promising alternative for environmental research and authorities.
Researchers from Drexel University have developed a machine-learning algorithm that can extract and recognize digital fingerprints of AI-generated video, including Stable Video Diffusion, Video-Crafter, and Cog-Video. The algorithm can be trained on just a few examples of new AI generators to detect them.
Researchers at DGIST developed a neural network module called DG-Net, which can accurately extract objects from aerial and satellite imagery. The technology has shown exceptional accuracy in geographic spatial object segmentation, outperforming existing models.
A novel approach combines LEACH clustering with fuzzy logic and ANN classifiers to detect intruders in wireless sensor networks (WSNs). The proposed method achieved high accuracy metrics, including 97% accuracy, precision, and sensitivity.
A research team has successfully created a new dimension in photonic machine learning by incorporating sound waves, enabling the creation of reconfigurable neuromorphic building blocks. This innovation has the potential to revolutionize computing tasks by providing high-speed and large-capacity solutions.
Researchers created a system called Holodeck to generate interactive 3D environments, leveraging language models like ChatGPT to control it. The system outperformed earlier tools in evaluating realism and accuracy, with human evaluators preferring its outputs across various indoor environments.
Researchers found that neural networks use a similar path to chart their way from ignorance to truth when presented with images, despite varying network designs and training recipes. This commonality holds the potential for developing more efficient image classification algorithms, reducing the computational power required by AI systems.
Researchers from Kobe University developed an AI image recognition algorithm that can predict mouse behavior based on brain functional imaging data, achieving 95% accuracy. The model identified critical cortical regions for behavioral classification and demonstrated near real-time speeds.
Scientists have developed a new method to manipulate light using synthetic dimension dynamics, enabling precise control over light propagation and confinement. This breakthrough has significant implications for applications such as mode lasing, quantum optics, and data transmission.
A new study highlights the importance of differentiating between formal and functional competence in language learning models. Researchers argue that leveraging human neuroscience insights can help develop more powerful AIs that mimic the brain's modularity, leading to improved performance and natural user interaction.
Researchers from UNIGE have developed an AI that can learn a task solely based on verbal instructions, then describe it to another AI, which reproduces the task. This breakthrough is promising for robotics and understanding human language.
Researchers at the University of California - San Diego developed a mathematical formula that reveals how neural networks learn to detect relevant patterns in data. The Average Gradient Outer Product (AGOP) formula helps interpret which features the network is using to make predictions, improving the accuracy and reliability of AI syst...
Politecnico researchers developed a new type of neural network called Latent Dynamics Network (LDNet) that can accurately predict the evolution of complex systems in low-dimensional spaces. This approach offers significant innovations over traditional methods, enabling up to 5 times more accurate results with a reduction of over 90% in...
A study using machine learning classifies galaxy mergers and finds that mergers are not strongly associated with black-hole growth. Cold gas at the center of the host galaxy is necessary for rapid growth, suggesting a more complex relationship between galaxy evolution and supermassive black holes.
The new AI model uses a visual map to explain each diagnosis, helping doctors follow its line of reasoning and check for accuracy. The tool aims to catch diseases in their earliest stages, making it easier on doctors and patients alike.
Scientists at NUS developed an AI-enabled atomic robotic probe to fabricate carbon-based quantum materials at the atomic scale. The CARP concept utilizes deep neural networks to autonomously synthesize open-shell magnetic nanographenes with precise engineering of their π-electron topology and spin configurations.
Explainable AI methods have been developed to make audio models more interpretable and transparent. Researchers categorize existing audio XAI methods into two groups: general methods and audio-specific methods, offering new possibilities for improving the trustworthiness of AI decision-making in audio tasks.
A new depth from focus/defocus approach, DDFS, combines model-based and learning-based strategies to achieve notable improvements in performance and applicability. The proposed method outperformed state-of-the-art methods in various metrics for several image datasets.
The study demonstrates that deep learning allows for precise observation of single molecules in artificial systems, cells, and small organisms. Deep learning-based methods reduce data requirements by an order of magnitude, enabling faster evaluation times.
Researchers developed an artificial neural network model to predict oxaliplatin-induced liver injury risk based on patient characteristics. The model outperformed traditional logistic regression in predicting the risk of liver injury, suggesting potential benefits for early screening and timely intervention.
The team proposed a novel machine learning model with data augmentation, which accurately predicts the plastic anisotropic properties of wrought Mg alloys. The model showed significantly better robustness and generalizability than other models, paving the way for improved design and manufacturing of metal products.
A neural network trained on a photographic database of olive fruit endocarps can identify olive varieties with high accuracy. The OliVaR tool automates the traditional morphological classification process, allowing growers to quickly identify olive tree varieties and advancing general knowledge of all existing olive varieties.
A new machine learning approach combines computer vision with deep-learning algorithms to pinpoint problem areas in concrete structures. The system enables efficient identification and inspection of cracks using autonomous robots, reducing the overall inspection workload.
Researchers have developed a novel optical neural network architecture that achieves nonlinear optical computation by precisely controlling ultrashort pulse propagation in multimode fibers. This approach streamlines the need for energy-intensive digital processes, achieving comparable accuracy with significantly reduced parameters.
The TRAILS AI Institute has awarded eight seed grants totaling $1.5 million to advance AI design, development and governance. The funded projects include developing AI chatbots for smoking cessation and designing animal-like robots for autism support.
A KAIST research team led by Professor Hawoong Jung identified the principle behind musical instincts emerging from the human brain without special learning using an artificial neural network model. The study found that cognitive functions for music forms spontaneously as a result of processing auditory information received from nature.
Scientists at Kyushu University use machine learning to identify promising green energy materials, accelerating the search for hydrogen fuel cell efficiency and expanding material discovery capabilities. Two new candidate materials with unique crystal structures have been successfully synthesized.
A team of South Korean scientists used machine learning to discover the secrets of cell variability, revealing a parallel structure that reduces heterogeneity among cells. This finding has far-reaching effects on cancer treatment and improvement in chemotherapy efficacy.
A recent study using AI to analyze registry data on people's residence, education, income, health, and working conditions can predict life events such as personality and time of death. The model outperforms other advanced neural networks and provides precise answers despite ethical concerns about sensitive data and bias.
A new study reveals AI tools are more vulnerable than thought to targeted attacks that force AI systems to make bad decisions. Researchers developed a software called QuadAttac K to test for vulnerabilities in deep neural networks.
This review article surveys existing deep active learning approaches, applications, and challenges in the context of foundation models. Effective query strategies and model training methods are essential for optimizing joint performance.
The European Union's AI act could enable AI to access our subconscious minds, potentially leading to manipulation. According to Ignasi Beltran de Heredia, only 5% of brain activity is conscious, and the remaining 95% operates subconsciously, making it difficult for us to control or even be aware of.
Researchers created an AI system with physical constraints to develop brain-like features, such as hubs and flexible coding schemes. The system solved a maze navigation task by combining multiple pieces of information, similar to human brains. This study sheds light on how physical constraints shape differences in brain organization.
Researchers at Rice University are developing a machine learning framework to improve decision-making processes in military communication networks. The goal is to enable rapid, adaptive action across a broad range of scenarios by combining local data in the most effective manner.
The Python code library snnTorch, developed by UC Santa Cruz's Jason Eshraghian, has surpassed 100,000 downloads and is used in various projects. A new paper published in the Proceedings of the IEEE documents the library and offers a candid educational resource for students and programmers interested in brain-inspired AI.
A new training algorithm enabled the brain-like system to learn and improve continuously, achieving higher accuracy than conventional machine-learning approaches. The system's ability to process multiple streams of data simultaneously makes it promising for tasks like pattern recognition in complex data.
A recent study published in the Proceedings of the National Academy of Sciences found that AI's deep convolutional neural networks can identify faces but struggle to capture other important information like emotional state and trustworthiness. Brain activity scans revealed a weak correlation between AI's codes and human brain represent...
The UTSA MATRIX AI Consortium has received a $2 million grant to create new AI models that rapidly learn, adapt, and operate in uncertain conditions. The team aims to bridge the gap between human brain processing efficiency and current AI limitations, enabling more efficient and adaptive AI systems.
A Lancaster University academic argues that AI and algorithms contribute to polarization, radicalism, and political violence, posing a threat to national security. The paper examines how AI has been securitized throughout its history, highlighting the need for better understanding and management of its risks.
Researchers are combining biology, physics, computer science, and engineering to design electric circuits that mimic the brain's adaptive behavior. The goal is to create a more efficient AI application that can learn from history and adapt without significant energy consumption.
Researchers at the University of Sydney have developed a physical neural network that can learn and remember data in real-time, using nanowire networks to mimic brain-inspired learning and memory functions. The network achieved high accuracy in benchmark image recognition tasks and demonstrated its capacity for online learning.
Researchers found self-supervised models generate activity patterns similar to mammalian brains, suggesting an organizing principle. The models learn representations of the physical world to make accurate predictions, potentially unlocking human-labeled data limitations.
The new AI chip uses ferroelectric transistors to store data and perform calculations, achieving a TOPS/W ratio of 885, twice as powerful as comparable chips. The goal is to use the chip for real-time applications such as deep learning and robotics, but security requirements and industry-specific criteria may delay its adoption.
Researchers developed an AI tool for automatic colorectal cancer tissue analysis, refining neural networks to offer faster and more precise classification. The tool's public accessibility aims to spur further advancements and breakthroughs in colorectal cancer research.
Researchers developed a novel photonic processor with adaptive neural connectivity, allowing for the creation of complex artificial neural networks. The system utilizes waveguide-coupled phase-change material to create almost 8,400 optical neurons that can adapt their connections through synaptic and structural plasticity.
Researchers developed a neural network model to accelerate detection of brain rhythms, reducing feedback signal delay from 500ms to 10ms. The approach shows promise for treating attention deficit disorder and epilepsy.
Researchers at Linköping University developed an AI-based method applicable to various medical and biological issues, accurately estimating people's chronological age and determining smoking status. The models identify previously known epigenetic markers used in other models, but also new markers associated with conditions.
Australian researchers developed an algorithm that can intercept and shut down man-in-the-middle cyberattacks on unmanned military robots. The algorithm, tested on a replica of a US army combat ground vehicle, was 99% successful in preventing malicious attacks.
Scientists at the Max Planck Institute present a method for training artificial intelligence using physical processes, reducing energy consumption and computing time. The new approach relies on non-linear processes, such as optics, to mimic the human brain's parallel processing, potentially leading to more efficient neural networks.
Researchers discovered that brain nerve networks are organized into interconnected modules to segregate and integrate inputs, enabling efficient processing. This modular architecture allows the brain to balance local activity with global integration, essential for information representation.
Researchers designed a simplified Mach-Zehnder interferometer mesh for real-valued matrix-vector multiplication, reducing hardware requirements and energy consumption. The new mesh detects incoherent light and is scalable, making it suitable for large-scale optical neural networks.