Articles tagged with Deep Learning
Keren Zhou receives $274,265 from NSF for a project on DLToolkit, a novel performance profiling and analysis infrastructure for scientific deep learning workloads. The toolkit aims to foster innovation in scientific applications using DL.
Researchers developed an innovative AI approach called GraSSCoL to predict complex astrochemical reactions. The model achieved outstanding Top-k accuracy scores, outperforming earlier state-of-the-art models by a significant margin.
A new AI tool can learn to read medical images with far less data, cutting down the amount of required data by up to 20 times. The tool improves upon medical image segmentation, a labor-intensive task often performed by experts, and boosts model performance in settings with limited annotated data.
A deep learning-based model enables fast and accurate stroke risk prediction by segmenting carotid arterial vessel lumens, vessel walls, and plaques in MRI images. The model achieves high accuracy in plaque segmentation, outperforming manual methods, and completes assessment in under 3 seconds.
A team of computer scientists created 2,300 original sudoku puzzles and asked AI tools like OpenAI's ChatGPT to solve them. The results showed that while some AI models could solve easy sudokus, most struggled to provide accurate explanations, raising questions about the trustworthiness of AI-generated information.
Researchers use AI to solve differential equations, such as Schrodinger's equation, for large-scale systems, improving efficiency and accuracy in fields like drug discovery and material design.
Researchers found that ChatGPT-4 performed better across demographic groups, while LLaVA showed significant sex-related biases in diagnosing skin diseases from medical images. The study emphasizes the need to address these biases to ensure AI models are safe and effective for all patients.
The team aims to deliver AI power directly to devices, improving resilience and speed in constrained environments. By processing data step-by-step across a network of devices, they can create a safe and adaptable system that can withstand attacks and extreme conditions.
Researchers developed an AI-based screening tool using pangrams to detect Parkinson’s disease with nearly 86 percent accuracy. The web-based test analyzes voice recordings for subtle patterns linked to the neurodegenerative disease, identifying potential warning signs.
Researchers developed a smart neural network model that combines CNNs and RNNs to predict multicolor soliton evolution, surpassing limitations of standard frameworks. The dual-channel system accurately tracks changes in energy, wavelength, and phase with remarkable accuracy.
Researchers developed an AI model that accurately differentiates patients with various neurodegenerative disorders using 3D brain imaging data. The model achieved high accuracy, particularly for vascular dementia and Alzheimer's disease, by focusing on subcortical brain structures.
Researchers found that deep neural networks exhibit absorbing phase transitions, a phenomenon observed in physical systems like forest fires. This discovery provides a unified framework describing how the signal propagates between layers of neurons, enabling prediction of trainability and generalizability.
Researchers developed OmicsTweezer, a tool that uses machine learning and single-cell data integration to analyze human tissue. The tool can estimate cell type composition in tumors and surrounding tissues, which could help pinpoint potential therapeutic targets.
A new deep learning model enhances handheld 3D medical imaging by automatically tracking transducer motion without external sensors. The model produces more realistic 3D US images and can reconstruct blood vessel structures using ultrasound and photoacoustic data.
Researchers developed a simple model that reproduces deep neural network features, allowing for optimized parameter tuning. The 'folding ruler' model demonstrates how nonlinearity and noise improve network performance, enabling more efficient training without trial-and-error.
Researchers developed an AI system that enables a four-legged robot to adapt its gait to different terrain, just like animals. The robot learned to switch gaits on the fly and navigate uneven surfaces without any alterations to the system itself, overcoming previous limitations around adaptability.
Researchers have developed an AI-assisted diagnostic system that can estimate bone mineral density in the lumbar spine and femur with high sensitivity and specificity. The system has the potential to transform routine clinical X-rays into a powerful tool for opportunistic screening, enabling earlier detection of osteoporosis.
Researchers developed an AI-powered diagnostic approach using quantitative biomarkers and biometrics to rapidly assess neurodivergent disorders. The method has the potential to diagnose autism or ADHD in as little as 15 minutes, providing healthcare providers with additional tools to tailor treatments.
A new technique called WeGeFT improves large language model performance in tasks such as commonsense reasoning and code generation. By fine-tuning key parameters, researchers reduce the need for significant computational power, advancing the field of artificial intelligence.
A new study reveals that AI systems transition from relying on word positions to meaning-based understanding as they receive enough data for training. The transition occurs abruptly, similar to a phase transition in physical systems, and is driven by the amount of data available.
The College of Engineering at Texas A&M is developing a suite of university-wide resources to integrate generative AI into course material, research, and outreach. The initiative aims to make generative AI a core part of the academic toolkit accessible to faculty across disciplines.
A team of researchers at Tohoku University's AIMR used machine learning potential to characterize Sn catalyst activity, identifying the most effective catalysts for CO2 reduction. The study provides novel insights into the behavior of Sn-based catalysts and could lead to more efficient fuel production.
A new study in Nature Communications found that AI models exhibit a geometric property called convexity, which helps humans form and share concepts. Convexity is also linked to the performance of AI models on specific tasks.
Researchers used machine learning to simulate galaxy evolution and supernova explosions, achieving speeds four times faster than supercomputers. This breakthrough enables the study of galaxy origins, including the creation of the Milky Way's elements essential for life.
The CNOP-DL method extends classical CNOP for deep learning methods, breaking deterministic causality and attributing forecast errors to all input slices. This new structure identifies critical time steps and locations where additional observations can significantly improve forecasts.
A research team developed an AI-based classification system using InceptionResNetV2 and DenseNet121 to identify five types of facial pigmented lesions. The system demonstrated high diagnostic accuracies compared to board-certified and non-certified dermatologists, with potential as a diagnostic support tool for clinical practice.
A review of AI eye imaging devices approved for patient care found significant gaps in evidence, including lack of transparency on training data and limited diversity in clinical evaluations. The study highlights the importance of rigorous, transparent evidence and data to ensure equitable and effective AI-based solutions.
Researchers have developed a new method to physically restore original paintings using digitally constructed films that can be removed if desired. The process uses a polymer film mask printed on a very thin film and aligned to an original painting, which takes around 3.5 hours from start to finish.
Researchers used proactive and transfer learning strategies to mitigate data shifts in AI models for hospital applications. They found that models trained on one hospital type performed better than those trained on all hospitals using transfer learning.
Assistant professor of electrical and computer engineering Natasa Miskov-Zivanov is receiving a $581,503 NSF CAREER Award for her project that leverages AI to design more effective lymphocytes for cancer immunotherapies. The system aims to accelerate the process of designing new therapeutic cell designs.
A refined artificial intelligence (AI) tool has shown promise for objective evaluation of patients with facial palsy. The 'fine-tuned' model demonstrated substantially lower error rates and improved keypoint detection in every area of the face, including areas of asymmetry.
A new near-real-time prediction model for earthquake-triggered landslides has been developed, utilizing a global database of 398,698 mapped events and cutting-edge deep learning. The model achieves spatial accuracy exceeding 82% and can generate probability maps of landslide occurrence in under one minute.
A new deep learning model, ENDNet, significantly enhances subgraph matching accuracy by identifying and neutralizing extra nodes that interfere with the matching process. This improves performance in pattern recognition tasks across various fields, including drug discovery and natural language processing.
A new microscopy method, LICONN, developed by ISTA scientists and Google Research, can reconstruct mammalian brain tissue with all synaptic connections between neurons. This technique uses standard light microscopes and hydrogel to achieve high resolution and opens up possibilities for visualizing complex molecular machinery.
A new study from the University of Surrey explores how metaverse platforms like Roblox and ZEPETO are changing consumer engagement forever. Digital doppelgängers, using AR and VR technologies, create immersive experiences that enhance brand engagement and emotional connections.
A portable device can instantly detect dangerous street drugs at extremely low concentrations, highlighting their dangers. The device, being trialled by drug-checking services in the UK, Norway, and New Zealand, allows for cheap and on-the-spot analysis of substances.
Researchers developed a novel vote-based model for accurate hand-held object pose estimation, addressing issues with existing approaches. The new framework achieves significant improvements in accuracy and robustness, enabling robots to handle complex objects and advancing AR technologies.
A new review advocates for building confidence in AI applications by implementing robust data governance frameworks, enhancing transparency, and involving stakeholders. The authors emphasize the importance of addressing ethical implications and ensuring equitable access to AI-driven innovations in clinical oncology.
Recent high-quality deepfake videos can feature realistic heartbeats and minute changes in face color, making them challenging to detect. Researchers found that even small variations in skin tone and facial motion can replicate the original pulse in deepfakes.
Researchers from OIST and Hebrew University developed a novel method to measure energy usage during movement using video and 3D-tracking via deep learning. This innovative approach expands the study of movement energy in ecology, physiology, and beyond, enabling the accurate measurement of energy consumption in smaller animal species.
A new AI tool has been developed to predict the relapse of pediatric brain cancer with high accuracy, using temporal learning algorithms to analyze sequential brain scans. The tool achieved an accuracy of 75-89% in predicting recurrence, outperforming predictions based on single images.
Researchers developed an IEAC framework combining robust security with high-capacity transmission performance, achieving a record 1 Tb/s secure transmission over 1,200 km of optical fibre. The system eliminates the trade-off between security and speed by integrating encryption into the communication process.
A Lehigh University team developed a novel machine learning method to predict abnormal grain growth in materials, enabling the creation of stronger, more reliable materials. The model successfully predicted abnormal grain growth in 86% of cases, with predictions made up to 20% of the material's lifetime.
Dr. Latifur Khan, a renowned computer science professor, has been elected as an AAAS Fellow for his pioneering work in machine learning and big-data analytics. He developed innovative solutions to adapt machine learning models to cybersecurity risks and created an AI-driven tool to analyze political conflict and violence.
A new hardware platform for AI accelerators capable of handling significant workloads with reduced energy requirement has been developed. The platform leverages III-V compound semiconductors to create photonic integrated circuits, which operate at the speed of light with minimal energy loss.
Researchers developed a method to detect epileptic seizures in humans using canine EEG data. The approach leverages feature similarities across species and modalities, reducing input space discrepancies. Euclidean alignment and knowledge distillation are key components of the proposed joint alignment mechanism.
Researchers at Concordia University have developed a new approach to identifying fake news on social media using the SmoothDetector model. The model integrates probabilistic algorithms with deep neural networks to capture uncertainties and patterns in multimodal data, providing more nuanced judgments of authenticity.
Researchers have introduced Orion, a novel framework that brings fully homomorphic encryption to deep learning, enabling computations on encrypted data without decrypting it. The framework achieves a 2.38x speedup over existing state-of-the-art methods and enables high-resolution FHE object detection using large neural networks.
Researchers developed an AI model that classifies variable stars from light curves with high accuracy, outperforming traditional approaches. The StarWhisper LightCurve series achieves near 90% accuracy with minimal manual intervention, paving the way for parallel data analysis and multi-modal AI applications in astronomy.
The study systematically traces Generative AI evolution from deep learning to foundation models, highlighting four distinct stages and successful applications. Key challenges like safety concerns and theoretical breakthroughs require further attention and development in the field of Generative AI.
A research team at Kumamoto University developed a deep learning-based method for analyzing the cytoskeleton more accurately and efficiently than ever before. This technique enabled more reliable measurements of cytoskeleton density, which is critical for understanding cellular structure and function.
Researchers from Osaka Metropolitan University used a deep learning model to discover new bubble-like structures in the Milky Way galaxy, providing insights into star formation and galaxy evolution. The study also revealed shell-like structures formed by supernova explosions.
A new AI model developed by UC Riverside scholars combines historical sales data with economic demand theory to predict prices in uncertain times. The model retains high accuracy, demonstrating a substantial improvement over other methods in reducing generalization errors.
Researchers developed a deep learning algorithm to denoise ultra-low dose CT scans, improving image quality and accuracy. The study found that this approach can diagnose pneumonia in immunocompromised patients using only 2% of the radiation dose of standard CT scans.
Researchers developed single-shot super-resolved fringe projection profilometry (SSSR-FPP) using deep learning to achieve 100,000 frames-per-second 3D imaging. This breakthrough offers new insights into ultra-fast dynamic processes and could revolutionize fields like mechanics and biology.
A new study suggests that artificial intelligence can effectively detect wildfires in the Amazon rainforest, using satellite imaging and deep learning. The technology achieved a 93% success rate in training models via datasets of images with and without wildfires.
This study utilized deep learning models to diagnose and predict the likelihood of malignant transformation in oral potentially malignant disorders. AI-driven approaches offer noninvasive, cost-effective, and objective means to enhance early detection and improve patient outcomes.
A deep learning model, CGMformer, leverages large-scale continuous glucose monitoring (CGM) data to extract individual glucose dynamics. The model captures a continuous picture of glucose fluctuations, identifying patterns that may indicate early metabolic dysfunction.
A new deep learning approach called HikingTTE significantly improves hiking travel time estimation by considering individual walking ability and fatigue. The model outperformed conventional techniques, reducing Mean Absolute Percentage Error by 12.95 percentage points.
The National Center for Supercomputing Applications has been awarded a grant to continue its Research Experience for Undergraduates (REU) program, which provides students with hands-on experience in machine learning and deep learning projects. The program aims to develop open source machine learning models and tools and apply them to r...