Research using machine learning to solve scientific problems. Covers AI quality control, sea-surface temperature estimation, label-free cell analysis, and cyber defense.
Researchers found that large language models exhibit systematic biases when evaluating texts, but only when the source or author is revealed. The studies showed a high level of agreement among models when no information was provided, but decreased agreement and even bias emerged when fictional sources were used.
TorchSim, a PyTorch-based simulation engine, delivers acceleration for MLIPs by unifying molecular dynamics and gradient-based learning. The platform provides speed, flexibility, and ease of integration with emerging machine learning atomistic models.
KVzip reduces chatbot response time and memory cost while maintaining accuracy, achieving 3–4× memory reduction and approximately 2× faster response times. The technology also demonstrates scalability to extremely long contexts and has been integrated into NVIDIA's open-source library.
A new AI model uses machine learning to predict drug toxicity in humans by identifying biological differences between cells, mice, and humans. The model improved predictive power over existing state-of-the-art models and demonstrated practicality in predicting market withdrawal due to toxicity.
A Michigan State University-led study examines how well AI personas can detect human deception and compares their performance to humans. The results show that AI is more lie-biased and less accurate than humans, highlighting the need for improvement before generative AI can be used for deception detection.
The FAU College of Engineering and Computer Science has established the 'Ubicquia Innovation Center for Intelligent Infrastructure' to develop transformative technologies. The center will empower students and faculty to create AI-First solutions for a smarter, more connected world.
The new tool, called FSNet, combines machine learning and optimization to find feasible solutions quickly while ensuring constraints are met. It can unravel complex problems several times faster than traditional solvers and even outperform pure machine learning approaches.
A UT Dallas team developed an electrochemical biosensor that accurately identifies eight volatile organic compounds linked to thoracic cancers. AI analysis enhances the accuracy of breath samples, showing promise for early lung cancer detection and improved patient outcomes.
Autograph, a new framework, uses graph neural networks and deep reinforcement learning to achieve higher accuracy and faster execution of compute-intensive programs. It outperformed other approaches across various datasets, with notable improvements on Polybench, NPB, and SPEC 2006 benchmarks.
Researchers Prof Axel Cleeremans, Prof Anil Seth, and Prof Liad Mudrik warn that advances in AI and neurotechnology are outpacing our understanding of consciousness. They emphasize the need for theory-driven research and innovative methods to advance consciousness science.
Researchers developed a machine learning technique to analyze plasma emission spectra, accurately identifying valence states and predicting film growth rates. This method uses full-wavelength information and can be used for real-time film deposition control technology.
A new collaboration between Rice University and Baylor College of Medicine aims to predict kidney injury sooner after heart surgery, potentially reducing hospital stays and mortality. The project uses artificial intelligence to analyze vast amounts of data from thousands of cardiac surgery patients.
Researchers warn that advances in AI and neurotechnology are outpacing our understanding of consciousness, with potential serious ethical consequences. A better understanding of consciousness could have major implications for AI, prenatal policy, animal welfare, medicine, mental health, law, and emerging neurotechnologies.
Researchers have identified replication protein A (RPA) as an essential protein for maintaining telomeres, which are protective caps at the ends of chromosomes. This discovery has significant implications for understanding and treating diseases caused by shortened telomeres, such as aplastic anemia and acute myeloid leukemia.
Researchers found that AI models predict protein structures despite modifications in amino acid sequences or ligands, indicating a lack of understanding of physical chemistry. The models only recognize patterns they've seen before and struggle with unknown proteins.
Researchers at USC Viterbi School of Engineering have developed artificial neurons that physically embody the analog dynamics of biological brain cells. These innovations will allow for significant reduction in chip size and energy consumption, potentially advancing artificial general intelligence.
Researchers developed a machine learning-based workflow, SPaDe-CSP, to predict crystal structures of organic molecules. The workflow narrows the search space by predicting probable space groups and crystal densities before computationally intensive relaxation steps.
AI algorithms can empower workers when used as tools for collaboration, rather than control. When managers use algorithms to explain system logic and give staff power to question decisions, employee dignity is protected.
A research team provides a framework to support doctors in their patient care while ensuring AI doesn't undermine their expertise. The framework addresses key issues like timing, trust, and over-reliance on AI.
Researchers examine how genAI affects task structure, worker adoption, and job displacement. The authors suggest genAI will widen the 'cone of automation,' substituting for complex work and infrequent tasks.
Pusan National University researchers develop a novel prompting technique to improve ChatGPT's accuracy in predicting fashion trends. The study reveals that ChatGPT can capture emerging themes and identify new trends not found in existing data.
Researchers at Science Tokyo have provided the first mathematical proof that reentrance implies temperature chaos in spin glasses. The breakthrough enhances understanding of disordered systems and has potential applications in machine learning and quantum technologies.
The Stowers Institute has appointed its first AI Fellow, Sumner Magruder, to harness the potential of artificial intelligence in biological research. He will collaborate with researchers to design new algorithms and unlock insights from large datasets.
Researchers develop AI-powered methods for modeling the Gulf of Mexico's dynamics, achieving higher accuracy for short-term predictions and emulating 10-year dynamics without hallucinations. This breakthrough drives forward critical management of natural resources in the U.S. and Mexico, advancing AI technology in earth sciences.
A recent study from Harvard John A. Paulson School of Engineering and Applied Sciences uses wearable sensor technology and machine learning to estimate ground-reaction forces in runners. This data can provide insights into performance and injury, enabling the development of devices that deliver real-time feedback to users.
A study published in the British Journal of Health Psychology reveals that commercial fitness apps can have negative themes such as quantifying diet and physical activity challenges, oversimplified algorithms, and aversive emotional responses. The findings suggest a need for user-centered design prioritizing wellbeing over rigid goals.
Researchers at HUN-REN Szegedi Biológiai Kutatóközpont have developed an AI-powered platform for automated 3D cell culture analysis, enabling high-precision screening of cellular models. The technology removes the limitation of throughput in personalized medicine, allowing for fast and accurate analysis of clinical samples.
A machine learning model developed by Dr. Lan Mu's team at Tianjin University of Commerce predicts biochar yield and nutrient content with stunning accuracy, unlocking smart soil solutions for healthier soils, cleaner ecosystems, and smarter farming.
Researchers develop geophysical-machine learning tool that estimates soil strength parameters using limited borehole data, enabling continuous subsurface characterization. The approach reduces the need for expensive and time-consuming drilling in challenging terrains.
The team aims to build an anonymized database representative of the whole population by collecting two vivid memories from participants. The findings will inform new ways to help people remember in more detail and understand human memories across the lifespan.
A new study from McGill University found that increased physical activity in older adults with cardiovascular conditions is linked to greater social support and access to greenspace. Brain imaging revealed a connection between brain connectivity and exercise behavior change.
Researchers at UC San Diego developed a new method for fine-tuning large language models with significantly less data and computing power. This approach updates only the necessary parts of the model, reducing costs and improving generalization.
Researchers developed voltage-matrix nanopore profiling to accurately classify proteins in complex mixtures based on their electrical signatures. The method reveals molecular individuality and compositional differences without labeling or modifications, holding promise for disease diagnosis and real-world bioanalytical applications.
A team of researchers at the University of Waterloo developed a framework that uses mathematical tools and machine learning to rigorously check and verify the safety of AI-driven systems. The framework has been tested on challenging control problems and matched or exceeded traditional approaches.
Researchers developed a machine learning-driven design for a high-energy NASICON cathode that surpasses previous materials in terms of specific capacity, average operating voltage, and rate capability. The new cathode addresses sustainability concerns by replacing toxic vanadium with more environmentally friendly elements.
Researchers at University of California San Diego have developed an AI-based method to target cancer stem cells, which can spread and resist therapy. The approach leverages machine learning to identify treatment targets and restore function to genes, leading to the self-destruction of cancer cells.
Lehigh University researchers used machine learning to compare bone marrow extracted from the hip and shoulder, finding six proteins that distinguish between the two extraction sites. This study may lead to standardized BMAC extraction protocols and personalized treatments based on protein concentrations.
Recent research found that large language models are not yet able to consistently fool humans in conversations. They struggle with using discourse markers, opening and closing features, and subtle imitations. Despite rapid development, key differences between human and artificial conversations will likely remain.
A landmark study analyzed health data from over 600,000 patients across 10 countries to assess patient risk for non-ST-elevation acute coronary syndrome (NSTE-ACS). The AI-powered model GRACE 3.0 predicts risk more accurately and guides personalized treatment decisions.
A new large language model, LassoESM, has been developed to predict lasso peptide properties, enabling the acceleration of rational design for biomedical applications. The model was trained on thousands of lasso peptide sequences and demonstrated accurate prediction of various properties.
VFF-Net applies label-wise noise labelling, cosine similarity-based contrastive loss, and layer grouping to improve image classification performance compared to conventional forward-forward networks. The algorithm reduces test errors on various datasets, enabling lighter and more brain-like training methods that make AI more sustainable.
The conference explores how generative AI is reshaping kidney medicine through AI-driven diagnostics, data integration, and omics analysis. Key findings include the use of LLMs to transform diagnostic precision, patient management, and research design.
Researchers from MIT and the MIT-IBM Watson AI Lab have introduced a new training method that enables vision-language models to localize personalized objects in a scene. By using carefully prepared video-tracking data with contextual clues, the model is better able to identify the location of a specific object in a new image.
A new AI-powered tool, EZSpecificity, can predict the best enzyme-substrate combination for various applications. The tool outperformed existing models in accuracy, especially for halogenase enzymes.
Researchers developed an AI system, InfEHR, that links unconnected medical events over time, revealing diagnostic insights. The system transforms millions of scattered data points into actionable patient-specific diagnostic insights.
Scientists from Japan developed a theoretical framework that explains how collective cells can perform complex tasks. The key is distributed information processing and reinforcement learning in the environment.
A new AI tool, SpectroGen, uses generative AI to quickly assess material quality by generating spectra in less than one minute. It can replace traditional methods that take several hours or days, improving productivity and efficiency in industries such as manufacturing and pharmaceuticals.
A new project aims to enhance workforce readiness in molecular bioscience by creating open-access resources and modules tailored to student needs. The Molecular Data Education Hub will host instructional materials and case studies for instructors to implement into their courses.
MetaSeg achieves the same segmentation performance as U-Nets but requires 90% fewer parameters, making medical image segmentation more cost-effective. The new approach leverages implicit neural representations to quickly adjust to new images and decode accurate labels.
Researchers are developing a precision phage platform to restore microbiome balance and combat antibiotic-resistant diseases. The MIGHTY project aims to harness phages as targeted antimicrobials, leveraging AI and machine learning methods to identify effective phage combinations.
A new project aims to develop a computationally efficient model that accurately predicts how additive manufacturing process parameters influence the solidification microstructure of binary alloy solidification. This will enable optimization of additively manufactured parts with confidence in critical industries.
The platform allows users to create personalized LLM-based chatbots, serving as a 24/7 teaching assistant or scraping campus websites to find needed resources. Illinois Chat has already been utilized in various departments and research groups, providing valuable feedback to make it more practical.
A new smartphone-based AI system accurately predicts avocado firmness and internal quality with high accuracy, enabling consumers to avoid overripe avocados. The technology has the potential to assess the ripeness and quality of other foods, reducing global food waste by 50% by 2030.
A new machine learning framework integrates experimental features with microbial biofilm data to optimize bioelectrodechlorination, predicting pollutant degradation rates with high accuracy. The approach reduces reliance on exhaustive laboratory testing while enhancing remediation efficiency and can be adapted for other bioelectrochemi...
Researchers at MIT have found a hidden atomic order in metals that changes their properties, including mechanical strength and heat capacity. The discovery reveals a new physical phenomenon explaining the persistent patterns and provides a simple model to predict chemical patterns in metals.
Researchers developed an AI-based generative approach to discovering technology opportunities from patent maps using machine learning. The system translates patent vacancies into human-readable text, enabling the identification of untapped technologies and facilitating innovation forecasting.
Researchers at Texas A&M University have developed a new framework to protect industrial processing facilities from cyber threats. The framework identifies vulnerabilities, detects abnormal activity in real-time, and provides safeguards and mitigation strategies to maintain safe operations.
FastTrack uses machine learning force field and 3D PES sampling to predict atomic migration barriers accurately within minutes, overcoming limitations of traditional methods. It enables quantitative high-throughput screening of ion transport across vast material spaces.
A new study uses general-purpose AI to classify real cosmic events and explain its reasoning without complex training. The model achieved approximately 93% accuracy and provided plain-English explanations for every classification.
Researchers at MIT have developed a 3D-printable aluminum alloy that is five times stronger than traditionally manufactured versions. This breakthrough could lead to lighter and more efficient aircraft parts, such as fan blades in jet engines, reducing energy consumption and costs.