Research using machine learning to solve scientific problems. Covers AI quality control, sea-surface temperature estimation, label-free cell analysis, and cyber defense.
Researchers at MIT developed a new method that coaxes AI models to achieve better accuracy and clearer explanations in safety-critical applications. The approach extracts concepts the model has learned while training for a specific task and forces it to use those, producing better explanations than standard concept bottleneck models.
MIT researchers developed a generative AI-driven approach for planning long-term visual tasks, surpassing existing techniques with a 70% success rate. The system combines vision-language models with formal planners, enabling robots to navigate complex environments and assemble multi-robot teams with high efficiency.
Jeonbuk National University researchers have developed DDINet, a lightweight and scalable model that can accurately predict drug-drug interactions for new, unseen drugs. This approach avoids overfitting to training data and is designed to handle binary and multi-classification tasks.
New research published in Weather assesses the energy consumption of AI models and finds that they consume considerable energy during training, but offset this by rapid forecasting ability. AI data-driven models are estimated to consume at least 21 times less energy than traditional models over one-year usage.
Researchers developed a new noninvasive brain stimulation technique by combining focused ultrasound with electrical stimulation, producing stronger, targeted brain responses. This approach, called transcranial electro-acoustic stimulation, clarifies conflicting results in the field and introduces a new approach to noninvasive brain sti...
Researchers developed Zephyrus, an AI agent capable of analyzing and answering questions in natural language about weather and climate data. The agent can handle language-based queries, translating them into code and generating plain language answers.
A new study by Mount Sinai researchers found that distributed multi-agent AI systems can maintain superior accuracy levels while using significantly fewer computing resources compared to single-agent designs. This approach can help healthcare organizations scale AI without sacrificing quality or safety.
A Chinese Neurosurgical Journal study developed a radiomics-based machine learning model to identify high-risk patients with traumatic brain injury who require emergency decompressive surgery. The model accurately distinguished patients who later required secondary surgery, suggesting its potential to complement clinical judgment.
A major UK study found that local councils are progressing at varying speeds towards AI adoption, with some councils building robust data foundations and others struggling with legacy systems. The report highlights the importance of leadership ambition, governance discipline, and strategic clarity in determining AI readiness.
Engineers at the University of Pennsylvania have developed LIBRIS, an automated microfluidic platform capable of generating lipid nanoparticle formulations at high speed and scale. This enables the creation of large, systematic datasets needed to train predictive AI models, accelerating the design of lipid nanoparticles for mRNA delivery.
Scientists at St. Jude Children's Research Hospital developed BOUQUET to analyze 3D-enhancer architecture in machine learning-based graph theory framework, identifying protein condensates and predicting gene expression. The findings provide new insight into how cells regulate genes controlling specialized identities.
Scientists at the University of Sydney have developed an ultra-compact AI chip that harnesses the power of light to perform calculations, potentially lowering energy consumption and increasing speed. The prototype, built in-house, achieved 90-99% classification accuracy in image classification tasks.
The Ateneo Laboratory for Intelligent Visual Environments (ALIVE) is developing machine learning solutions with industry partners to improve public health, traffic systems, and more. By bridging the gap between messy reality and mathematical models, ALIVE is creating intelligent visual systems that can handle real-world conditions.
A new study suggests that over 40% of UK adults are happy to use ChatGPT for mental health support due to long waiting times for traditional services. However, experts caution that AI should not replace human healthcare professionals and raise concerns about the potential impact on education and physical health.
Researchers at Worcester Polytechnic Institute developed an AI tool that analyzes brain scans to predict Alzheimer's disease. The method is accurate in detecting the disease among normal brains and those with mild cognitive impairment, highlighting key anatomical changes such as volume loss in the hippocampus and entorhinal cortex.
Researchers create metallic glass with exceptional kinetic stability while retaining ductility, opening new avenues for high-performance amorphous materials. The discovery uses oxygen patterning to decouple properties, allowing for tailored material design with unprecedented precision.
A new study explores the market effects of unleashing generative AI on creative endeavors, finding that low-quality 'AI slop' harms consumers and professionals. High-quality AI, however, can enhance professional work while providing consumers with better content options.
A York University-led study used machine-learning models to analyze 64 immune biomarkers in people with and without HIV who received COVID-19 vaccines. The researchers found clear vaccine-initiated immune response biomarkers between the two groups, but also identified outliers that provide insights into the complex nature of the immune...
A new machine learning approach accelerates Raman spectrum prediction for fast-ion conductors, revealing liquid-like ion motion. The method identifies low-frequency Raman signatures associated with high ionic mobility and relaxational host-lattice dynamics.
A recent study found that AI programs are up to 97% accurate at detecting pictures of deepfake faces, but performed at chance levels when it comes to detecting deepfake videos. Humans correctly identified real and fake videos about two-thirds of the time, picking up on subtle inconsistencies in movement and facial expressions. The stud...
Scientists have identified five distinct species of antbirds, including two new species, by integrating AI, vocal analysis, and traditional museum work. The study reveals that populations separated by major Amazonian rivers have evolved into distinct species.
A new AI model, TweetyBERT, automatically segments and classifies canary vocalizations with expert-level accuracy, providing insights into the neural basis of human language. The model, developed by University of Oregon researchers, uses self-supervised machine learning to analyze birdsongs and identify communication units, potentially...
Recent advances in machine learning enable the non-targeted analysis of thousands of chemical features in a single environmental sample. Machine learning models can predict tandem mass spectra from known molecular structures and infer molecular formulas, significantly narrowing down candidate structures.
The journal argues that AI can process vast datasets to detect patterns and predict climate impacts, improving pollution tracking, climate modeling, and agricultural planning. AI-driven modeling can also help policymakers evaluate tradeoffs and anticipate unintended consequences, supporting more informed decision-making.
A team of researchers has developed DopFone, an app that uses a smartphone's speaker and microphone to estimate fetal heart rate with high accuracy. The system was tested on 23 pregnant women and showed promising results, with an average error of 2 beats per minute.
Researchers have developed a smaller and simpler AI model that accurately predicts neural responses to visual stimuli in macaque brains. The compact model reveals unique neuron preferences for features like edges and colors, shedding light on how the brain processes information.
The semiconductor industry is shifting from compute to memory as ultra-large AI models demand higher performing chips. SK hynix is increasing bandwidth by a factor of 1.5x every two years, while GlobalFoundries uses AI to improve process control and manage diverse manufacturing processes.
Researchers used machine learning techniques to compress a large model of the visual cortex, creating smaller versions that predict neural responses with high accuracy. The compact models revealed specific computational patterns in how neurons detect important features, offering insights into how visual information is processed.
A new study finds that fewer than 250 fossils are required to train an image-based AI algorithm, a significant improvement on previously thought numbers. The discovery could greatly speed up the identification process in vertebrate paleontology, where most fossils are fragmented and difficult to analyze.
A study reveals that identical photons in optical circuits exhibit Hopfield Network behavior, enabling associative memory mechanisms similar to the human brain. The research finds a fundamental limit to memory capacity, with quantum coherence allowing correct retrieval but transitioning to disorder as data volume increases.
Researchers at Linköping University have developed an AI model that can accurately determine the time of death from blood samples, providing crucial information in murder investigations and police work. The method uses metabolites to analyze changes in small molecules over time, outperforming current methods.
Researchers developed a machine learning-guided strategy to design advanced biochar materials that remove phosphorus efficiently while lowering treatment costs. The study provides a practical pathway for restoring eutrophic waters at large scale.
Researchers at Linköping University have developed an AI-boosted electronic nose that can detect ovarian cancer from blood plasma samples with high accuracy. The method uses machine learning to identify patterns specific to the disease, making it a promising tool for early detection and improved survival rates.
Epic Games has acquired Meshcapade, a Max Planck startup that develops solutions for creating and animating digital humans. The technology, based on the SMPL body model, enables realistic human movement and expression in 3D.
Researchers developed a new model to predict long-term survival after heart failure in elderly Japanese patients. The Top-20 XGBoost model incorporates physical function metrics, rivaling the importance of traditional cardiovascular risk factors, and provides a more accurate estimation of mortality risk.
A team from MIT and UC San Diego has developed a new method to uncover hidden biases, moods, and abstract concepts in large language models (LLMs). The approach identifies these connections within the model and allows for manipulation of the concept in generated answers.
A new machine learning model interprets leg motion as expended energy, providing a more accurate measure of calories burned. The device has been shown to have double the accuracy of commercial smartwatches and activity trackers.
A new study by the University of Cambridge found that many AI bots lack basic safety disclosures, including transparency about their abilities and potential risks. The 'AI Agent Index' revealed a significant transparency gap, with only four out of thirty agents publishing formal safety and evaluation documents.
A study found that patients value AI-based interpretation for routine situations due to its speed and convenience, while remote video interpretation is preferred for emotionally sensitive or high-stakes conversations. Patient trust, autonomy, and perceived control drive interpreter preferences.
Researchers found that dosed nonlinearity improves model performance in various tasks, especially with limited data. Nonlinear units function like flexible switches, adapting linear processing modes based on context.
Researchers at UCSF and Wayne State University found that generative AI tools can perform orders of magnitude faster than human teams in analyzing health data. Junior researchers paired with AI generated viable prediction models in minutes, outperforming experienced programmers in hours or days.
A study by RIKEN researchers identifies a MYCN-driven biomarker that predicts the risk of liver cancer. The biomarker, known as the MYCN niche score, uses machine learning to analyze gene expression patterns and indicates whether a tumor-free liver is at high risk for developing tumors.
The journal explores the convergence of computational biology, artificial intelligence, and healthcare innovation, with a focus on precision medicine and enhanced patient care. Submissions are accepted from researchers, clinicians, and technologists on topics such as AI in medicine, computational genomics, and drug discovery.
Researchers developed an AI-based tool called BIOPREVENT to identify patients at higher risk for chronic GVHD and dying from transplant-related causes. The tool combines immune biomarkers, clinical data, and machine learning to create a personalized risk estimate over time.
China's youth faces a growing mental health crisis, but AI platforms like DeepSeek offer promising solutions to bridge the gap. These platforms leverage natural language processing and generative AI to provide round-the-clock support tailored to Chinese society.
Researchers found that wild Yellow-naped Amazon parrots use complex vocalizations with syntax, collocates, and a large repertoire of notes when fighting for territory. The team identified over 450 calls in warble duets, revealing precise rules governing the birds' communication.
A University of Houston professor has found that tree-like thin films release heat at least three times better than traditional methods, enabling more efficient cooling in AI data centers. The discovery demonstrates the power of physics-aware AI design for validating high-impact cooling solutions.
Researchers developed an AI-based system to generate high-resolution soybean yield maps across Brazil, leveraging knowledge from U.S.-based models through transfer learning. The approach achieved strong predictive performance without using municipal-level yield data, improving estimates for this key agricultural region.
Rice University scientists create a detailed map of the Alzheimer's brain using hyperspectral Raman imaging and machine learning. The findings show that chemical changes are unevenly distributed across the brain and extend beyond amyloid plaques, revealing broader metabolic differences between healthy and diseased brains.
Scientists have identified specific patterns of brain chemical activity in honey bees that predict how quickly individual bees learn new associations. The findings may help explain why humans learn at different speeds and provide implications for understanding brain disorders.
Researchers at the University of Pennsylvania have developed HoloRadar, a system that enables robots to reconstruct hidden 3D spaces beyond their line of sight using radio waves processed by AI. This capability can improve safety and performance in driverless cars and cluttered indoor settings.
A new study published in The Lancet Respiratory Medicine found that prehospital emergency intubation of high-risk trauma patients improves 30-day survival by 10.3% and could save 170 lives each year in the UK. Prehospital intubation needs to be administered by an advanced critical care team.
Researchers at Nagoya University developed an AI system called YORU that recognizes animal behaviors with over 90% accuracy. The system combines real-time video capture with optogenetics to selectively target brain cells driving specific behaviors, offering a major breakthrough in social behavior studies.
Researchers at Oregon State University have developed a deep learning-based model for rapid bacterial contamination detection, eliminating misclassifications of food debris. The enhanced model can reliably detect bacteria in three hours and has the potential to prevent outbreaks and protect consumer health.
New machine-learning tools significantly improve phishing site detection accuracy, surpassing 95% precision/recall. The study evaluates ten classifiers across three public datasets using URL, domain, and content features.
A recent NSF grant will support the development of new diagnostics and predictive models for understanding self-competition and weak asymmetry in turbulent flows. The project aims to uncover hidden patterns that current models miss, leading to improved simulations in weather forecasting, climate modeling, and engineering design.
Drexel researchers create machine learning program that integrates qualitative and quantitative data to identify gentrification in Philadelphia neighborhoods. The program, trained with data from thousands of images and focus groups, accurately identifies new-build gentrification with 84% accuracy.
Researchers analyzed over 25,000 YouTube comments on wildlife videos to find only 2% calling for conservation efforts. Despite the low number, experts believe there is an opportunity for improved conservation messaging on social media.
A new study from George Mason University reveals that AI-driven antidepressant treatment can be less effective for African American patients due to the use of general population data. The study suggests that incorporating additional patient demographics, such as race and ethnicity, can improve the tool's effectiveness.
Researchers used machine learning to simulate gold nanoclusters' behavior under realistic conditions, revealing structural changes and polymer-like chain formation. The study demonstrates the accuracy of the machine learning potential in understanding nanomaterials' properties.