Research using machine learning to solve scientific problems. Covers AI quality control, sea-surface temperature estimation, label-free cell analysis, and cyber defense.
A new project aims to improve the performance of Graph Neural Networks (GNNs) by leveraging weak supervision and additional information. The research has potential applications in fraud detection, agriculture, and cancer diagnosis.
A new algorithm, FusionM4Net, has been developed to classify skin lesions with improved diagnostic accuracy. The algorithm uses a multi-stage data fusion process and outperforms previous state-of-the-art algorithms.
Researchers found animal-dispersed plant species' ability to adapt to climate change reduced by 60% due to bird and mammal losses. Global seed dispersal mapping revealed severe declines in temperate regions, with tropical areas at high risk if endangered species go extinct.
A new platform uses machine learning to design and build transformable, inflatable systems with potential applications in medicine, architecture, robotics, space travel, and more. The researchers used finite element simulations and neural networks to learn how to control the deformation of membranes when pressurized.
A new PET/CT artificial intelligence model has been developed to predict the risk of future heart attacks with improved accuracy. By combining information from PET and CT angiography, the model provides a more robust prediction of heart attack risk than clinical data alone.
Researchers at Beckman Institute have defined a mathematical framework for identifying hallucinations in biomedical images. This framework will enable researchers and radiologists to quantitatively assess their image reconstruction methods and prevent patient misdiagnosis.
A team of researchers at Washington University in St. Louis used machine learning to understand how locusts can consistently recognize smells despite environmental factors, finding that combining the activity of ON and OFF neurons provides a simple yet effective solution.
Researchers found that two commonly used atomic fingerprints, ACSF and SOAP, are insensitive to certain movements, leading to the failure of machine learning in resolving four-body interactions. This limitation affects the accuracy of reproducing these interactions with limited success.
The European Research Council has awarded €1.5 million in grants to three Saarbrücken-based researchers. Their projects focus on artificial intelligence and cybersecurity, with aims to develop fairer machine learning algorithms and secure computing methods.
Researchers at NIMS successfully fabricated high-performance neodymium magnets using machine learning, optimizing processing conditions with limited experimental data. By leveraging active learning and Bayesian optimization, they were able to achieve better magnetic properties than conventional sintered magnets.
Researchers used AI tools to analyze tissue samples from 439 patients with head and neck cancers, identifying those who could benefit from reduced radiation therapy. The study aims to improve treatment decisions for patients, reducing side effects and improving quality of life.
A recent study found that conventional metrics for detecting audio adversarial examples are unreliable and fail to accurately represent human perception. Researchers proposed a more robust evaluation method, but acknowledge the complexity of modeling auditory perception with mathematical metrics.
New experiments challenge conventional wisdom on neuronal refractory periods, discovering durations exceeding 20 milliseconds and sensitivity to input signal origin. These findings may hold the key to understanding degenerative diseases and advancing artificial intelligence-based applications.
A pilot study suggests that machine learning algorithms combining EKG and electronic health record data can more effectively screen for pulmonary embolisms than current tests. The fusion model was found to be 15-30% more effective at accurately identifying cases, especially severe ones.
Researchers develop system to image millions of brain cell connections in real-time, revealing key locations for learning and memory encoding. The new tool allows scientists to study synapse activity on a massive scale, with potential applications in understanding diseases such as Alzheimer's and autism.
An analysis of data from 16 US states found that wine and spirit sales increased during the early months of the pandemic, while beer sales declined. Visits to liquor stores increased, but those to bars and pubs decreased.
Researchers analyzed hundreds of thousands of secure email messages between doctors and patients to find that most doctors use language too complex for their patients' low health literacy. Effective communication can improve patient outcomes by tailoring electronic messages to match the complexity of the patient's language.
Researchers developed an AI model that can diagnose COVID-19 with high accuracy, using federated learning to preserve patient data privacy. The model was trained on over 9,000 CT scans from 23 hospitals in the UK and China, and validated against a panel of radiologists.
Researchers simulated an attack that falsified mammogram images, fooling both AI breast cancer diagnosis models and human radiologist experts. The study highlights the need to develop ways to make AI models more robust to adversarial attacks, which could lead to incorrect cancer diagnoses.
A team of UBC Okanagan researchers has developed a system to enhance interactions between humans and robots in industrial settings. The system uses artificial intelligence and machine learning to capture and analyze the environment, allowing robots to respond in a timely manner to ensure human safety.
Researchers at Chalmers University of Technology have developed an algorithm that learns optimal energy usage for electric delivery-vehicles. By focusing on overall energy usage instead of just distance travelled, the vehicles can reduce their energy consumption by up to 20% and minimize battery usage.
Researchers developed an algorithm to differentiate life-threatening gunshot events from non-life-threatening plastic bag explosion events. The study found that 75% of plastic bag pop sounds were misclassified as gunshot sounds, highlighting the need for a diverse dataset of similar sounds.
Researchers developed an AI technique to predict material properties using a small number of experiments, improving accuracy and facilitating digital transformation in materials development. The technique uses Bayesian optimization and incorporates measurement data into machine learning models.
Researchers at MIT and Google Brain developed a system that predicts how changing materials or designs will improve solar cell performance. The new simulator, called differentiable solar cell simulator, provides information on which changes will provide desired improvements, increasing the rate of discovery of new configurations.
DeepMind's neural network approach accurately describes electron interactions in chemical systems, overcoming long-standing challenges. The company's breakthrough enables researchers to explore material design, medicines, and catalysts at the nanoscale level.
A global community of hackers and threat modellers is needed to stress-test the harm potential of new AI products. Companies can harness techniques like red team hacking, audit trails, and bias bounties to prove their integrity and earn public trust. The industry faces a 'crisis of trust' if it doesn't adopt these measures.
A team from the University of Washington has developed a non-destructive 3D imaging method that can help doctors more accurately diagnose borderline cases of prostate cancer. The new approach uses 3D images to identify complex features in tissue samples, which can increase the likelihood of correctly predicting a cancer's aggressiveness.
The study leverages machine learning to tackle long-unsolvable problems in biological systems at the cellular level. By reducing data points, researchers can better analyze and model the impact of cells with high fidelity.
Researchers at the Max Planck Institute for Intelligent Systems developed an algorithm that analyzes gravitational wave data in seconds, rather than hours or months. The system, called DINGO, uses a deep neural network to infer properties of binary black-hole sources with high accuracy.
The University of California, Riverside, has been awarded a $980,000 grant from the Department of Energy to develop an AI-driven detector for the future Electron-Ion Collider. The team will use machine learning techniques to optimize detector design and achieve 'co-design,' a new concept in nuclear physics.
Carlos Ponce is studying the parts of the visual system that analyze shapes, using macaque monkeys as a model. He combines computational models with electrophysiology experiments to understand how neurons process visual information.
Researchers developed a new technology that tracks thousands of cells and determines the precise moment of death for any cell in the group. The approach was shown to work in rodent and human cells as well as within live zebrafish, and can be used to follow cells over weeks to months.
A new project led by University of Illinois researchers will develop machine learning models to predict the reactivity of thousands of organic contaminants in engineered and natural environments. This will help scientists better model pollutant fate and transport, leading to more accurate contaminant risk assessments.
A recent study uses machine learning to rapidly discover bacterial isolates with antifungal properties, identifying promising new compounds for crop protection. The approach analyzes thousands of microbial genomes at once, allowing researchers to identify novel beneficial microbes and bypass traditional screening tactics.
A new machine learning-based algorithm can predict stable material compounds much faster than traditional methods, opening up new avenues for research and discovery. The researchers identified several thousand potential new compounds using the computer, offering a promising breakthrough in materials science.
A study by the University of Toronto's Rotman School of Management found that predictive analytics can increase revenue by $500,000 to $1 million for manufacturers who invest in IT capital, educate their workforce, and implement high-efficiency manufacturing processes. The research team surveyed over 30,000 manufacturers and found that...
A team of researchers used deep learning to analyze cardiac images and identified genetic variations linked to aortic size. The findings may lead to the development of a polygenic score to identify individuals at high risk of aneurysm, as well as new drug targets for aortic enlargement.
Researchers aim to develop AI agents that reuse information, adapt quickly to new conditions and collaborate by sharing experiences. The goal is to enable machines to continually learn from their collective experiences and improve performance on novel and previous tasks.
Researchers used machine learning to identify patterns in knot theory and representation theory, suggesting new connections that mathematicians were able to prove. This collaboration demonstrates the potential of AI as a tool for guiding intuition in mathematical research.
A Michigan Tech-developed machine learning model uses probability to classify breast cancer shown in histopathology images and evaluate the uncertainty of its predictions. The model outperforms similar models and can measure uncertainty, promising time savings and referrals to human experts.
Computer scientists and mathematicians have used artificial intelligence to help prove or suggest new mathematical theorems in complex fields. The breakthrough uses DeepMind's AI processes to explore conjectures in mathematics, leading to a completely new theorem in knot theory.
A new computational method has been developed to accurately predict oxide reactions at high temperatures, even without experimental data. This approach combines quantum mechanics with machine learning to design clean carbon-neutral processes for steel production and metal recycling.
A research team developed an AI framework that analyzes protein interactions to predict effective and low-toxicity cancer drug combinations. The framework, GraphSynergy, outperforms conventional models in identifying synergistic combinations.
Researchers found that major COVID-19 models were wrong and not very useful in predicting the pandemic course. They used physics-informed neural networks (PINNs) to capture changing parameters and make predictions with improved accuracy.
A new artificial intelligence framework called TinNet combines machine-learning algorithms and theories to identify new catalysts for efficient energy production. By understanding how catalysts interact with different intermediates, researchers can design robust catalytic processes that improve daily life.
Researchers developed transformational machine learning (TML) to learn from multiple problems and improve performance while learning. TML out-performs current machine learning methods for drug design, accelerating the identification and production of new drugs.
Researchers have developed a new AI technique, MuSIC, which combines microscopy, biochemistry techniques and artificial intelligence to reveal approximately 70 components contained within a human kidney cell line, half of which had never been seen before.
Researchers used AI to optimize multiple properties of flow batteries, finding molecules that store a lot of energy and remain stable. The study uses quantum chemistry-guided multiobjective Bayesian optimization to identify promising candidates.
A KAUST team developed an improved method for detecting malicious intrusions using deep learning, achieving accuracy rates of up to 99% in simulations of different kinds of attacks. This stacked deep learning approach promises an effective defense against cyberattacks and could prevent outages in critical infrastructure.
Researchers develop a methodological framework to extract and monitor information from consumer reviews, providing actionable insights on product attributes and their benefits. The study also extends sentiment analysis by demonstrating hierarchical sentiment analysis, enabling managers to generate tailored dashboards and inform decisions.
A new 'image analysis pipeline' called TDAExplore gives scientists rapid insight into how cells are changed by disease, using a combination of microscopy, topology, and artificial intelligence. This approach can provide objective information on cell changes, such as the movement of proteins like actin, even with limited training data.
A new web-based application uses machine learning to predict lupus nephritis treatment response, considering various disease indicators. The tool may help physicians identify patients at risk of poor outcomes, enabling them to provide targeted care and preserve as much kidney function as possible.
A new study by USC researchers uses GANs to generate synthetic neurological data that can be fed into machine-learning algorithms to improve BCI usability. This approach improved BCI training speed by up to 20 times and enabled rapid adaptation to new subjects.
Researchers developed a technology that accurately detects lies by analyzing facial muscle contractions, achieving a success rate of 73%. The study identified two distinct groups of 'liars' based on cheek muscle and eyebrow activation, with potential implications for real-life deception detection.
Machine learning enables better understanding of climate-induced hazards, predicting floods and landslides with high accuracy. The technology combines diverse data sources to assess risk extent, considering both triggering hazards and socio-economic vulnerability.
The research team developed a technology for remotely assessing the condition of a building during an earthquake based on the readings from the building's seismometer. The new method uses CNN machine learning to quickly assess damage levels and determine if a building can continue to be used.
Researchers developed EDS-HAT, an AI-powered system combining machine learning and whole genome sequencing to detect clusters of similar infections in real-time. The system identified 99 clusters of infections and prevented potential transmissions in 65.7% of cases, saving the hospital $692,500.
Researchers from Okayama University developed an AI-powered image classifier to simplify and speed up the task of image analysis in cell biology. The system achieved high detection accuracy for mitotic cells in plant species, demonstrating its potential for non-experts to use.
A new study used machine learning to predict the zoonotic capacity of 5,400 mammal species, identifying those at high risk of transmitting SARS-CoV-2. The model, which combined data on biological traits with ACE2 receptor information, predicted 72% accuracy and identified numerous additional species with potential to transmit the virus.
International researchers used machine learning to forecast marsh establishment under various environmental conditions, revealing that controllable local factors are more important than global climate change. The study suggests smart management of tidal flats can counteract threats and strengthen wetlands.