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 machine learning algorithm confirms 50 new exoplanets, ranging from Neptunes to Earth-sized worlds, by distinguishing between real and false positives. The technique outperforms previous methods, can be automated, and improves with further training.
Researchers developed a machine learning model that analyzes molecular structure to predict enthalpy of formation with better accuracy than traditional approaches. The model's accuracy improves with more data, enabling the development of fully automated algorithms for predicting complex chemical phenomena.
A team of researchers at Tokyo Tech successfully used machine learning with an artificial neural network model to predict two key properties of self-assembled monolayers, enabling advanced material screening and design. This approach opens up new possibilities for the development of biomaterials with desired functions.
Researchers have developed two new strategies to automate the assessment of stored red blood cell quality, matching and surpassing expert analysis. Trained machines can analyze vast quantities of images to accurately predict RBC degradation, eliminating human error and improving consistency.
KDD 2020 features four keynote talks on meta-provenance, AI for intelligent financial services, state-space multi-taper time-frequency analysis, and computational epidemiology. The conference will also include 18 applied data science invited talks and 217 accepted research papers.
The NASA JPL team is using deep learning to develop software for future Mars rovers, which will enable them to travel farther and explore more of the planet. The team has been training machine learning models on the Maverick2 supercomputer and developing novel capabilities such as Drive-By Science and Energy-Optimal Autonomous Navigation.
Researchers have developed a new approach to machine learning combining ensemble methods and deep learning to diagnose cancer, predict viral attacks, and revolutionize molecular biology. This emerging field has the potential to transform bioinformatics and biomedical sciences.
Researchers developed an algorithm that allows robots to improve navigation systems by watching a human drive, enabling them to navigate more quickly and with fewer failures. This approach uses machine learning from demonstration, leveraging the expertise of soldiers in training environments.
A new study by Cornell researchers uses machine learning to assess the effectiveness of mathematical tools in predicting financial markets. The model can also predict future market movements, a task considered extraordinarily difficult due to markets' massive amounts of information and high volatility.
Researchers at Babylon have developed a new approach to AI diagnosis using causal machine learning, which enables the algorithm to consider alternative realities and improve accuracy for written test cases. This technology has the potential to augment the work of clinicians and drive better healthcare outcomes.
Researchers developed a sensitive and specific early warning system for predicting NEC using stool microbiome features combined with clinical and demographic information. The pilot study showed optimal performance from a gated attention-based multiple instance learning approach.
A new machine learning approach, federated learning, enables hospitals to share patient data securely, improving the accuracy of brain tumor diagnoses. This technique has been successfully tested in a Penn Medicine study, which showed that it can outperform centralized models in identifying brain tumors.
The ERC Proof of Concept project 'AssemblySkills' aims to validate an autonomous, intelligent skill learning system for robots to acquire and improve motor skills. The project builds on the ERC Starting Grant 'SKILLS4ROBOTS', which has yielded a structured control architecture that can scale robot learning to complex tasks.
A team of researchers at UC Riverside used machine learning to understand what chemicals smell like, predicting how any chemical will smell to humans. This breakthrough technology has vast applications in the food, flavor, and fragrance industries, including discovering new flavors and insect repellents.
A new approach uses photons to perform computations required by neural networks, improving speed and efficiency. Photonic tensor cores can process data in parallel, reducing power consumption and increasing throughput.
Researchers at Graz University of Technology developed a new machine learning algorithm called e-prop, which significantly expands the possible applications of AI. This novel approach uses spikes to enable more efficient information processing and reduces energy consumption.
Researchers used 2018 Japan floods to calibrate a machine learning model to predict Typhoon Hagibis' flooding impact. The model accurately identified inundated areas, verifying AI's potential for learning from past disasters.
Researchers at MDC developed Janggu, a universal programming tool converting genomics data into a format compatible with deep learning models. This allows for flexible and efficient analysis of large datasets, enabling the investigation of various biological questions.
Zhi Tian will receive funding to develop communication-efficient approaches for collaborative learning from private data in big data computing. Her goal is to minimize overall runtime, communication costs and total samples used.
The Argonne team has created a machine learning algorithm that approximates how the present detector would respond to the greatly increased data expected with the LHC upgrade. This algorithm simulates detector responses and reconstructs objects from physical processes, enabling faster and more accurate analysis of particle collisions.
Researchers at Cornell University used AI to investigate how reflection changes images, discovering clues like facial features and beards that can differentiate originals from reflections. The study has implications for training machine learning models and detecting faked images.
A new study by UCLA professors combines AI and folklore analysis to examine the storytelling elements of debunked conspiracy theories and actual news stories. The researchers found that conspiracy theories tend to form around specific elements that act as an adhesive, holding facts and characters together.
The researchers aim to create a machine-learning solution that combines pure data-driven reinforcement learning algorithms with domain knowledge. Their technique, Dino-RL, is designed to dynamically adjust to quickly detect and fix network problems without human intervention.
The increasing use of machine learning systems in public policy raises concerns about accountability, bias, and transparency. Explaining complex algorithms can reveal conflicting aims and implicit trade-offs in policy decisions.
Researchers found that humans can complement machine learning in correcting for biases. Vintage-specific skills and domain expertise are key attributes that help humans guide machines in mitigating bias. Human collaboration improves ML productivity but its impact on long-term productivity is unclear.
A machine-learning model analyzes histopathology slides to predict cancer prognosis and survival. The study, published in PLOS ONE, shows promise for improving cancer care.
A team of researchers from the University of Konstanz and Innsbruck explore the role of artificial intelligence in basic research, focusing on agency, creativity, and authorship. They aim to provide a conceptual framework for the development of AI methods in science.
Researchers developed an AI algorithm that uses the 'slowness principle' to estimate age and ethnicity by ignoring rapidly changing facial features. The system achieves impressive accuracy, outperforming even human experts in face recognition.
Researchers have improved Wi-Fi network operation and performance for the 5G/6G ecosystem by applying machine learning techniques. A new algorithm, ε-sticky, is proposed to reduce service disruptions and network instability, benefiting both stations that have found a suitable access point and those that haven't.
Artificial neural networks became unstable after continuous unsupervised learning, but exposure to Gaussian noise mimics slow-wave sleep stabilized them. This finding has implications for the development of biologically realistic AI systems.
A study by Oregon State University found that telepresence robots can enhance student engagement and expression in online classes. In contrast, instructors preferred teaching students in person, but valued telepresence robots as a remote learning solution over distance learning tools.
Researchers developed a novel AI-managed trading strategy that outperforms traditional methods, achieving greater gains and fewer losses. The proposed system utilizes convolutional neural networks to analyze layered images of current and past market data, leading to more accurate predictions and reduced randomness.
A new study by Carnegie Mellon University researchers has found the brain programs that code the sequence of steps in performing a complex procedure. The main findings were that each knot had a distinctive neural signature, so the researchers could tell which knot was being tied from the sequence of brain images collected.
A new NIST formula could significantly improve Wi-Fi and cellular system performance in unlicensed bands by selecting optimal frequency channels. The formula uses machine learning to maximize data rates and reduce interference, with computer simulations showing it can outperform exhaustive trial-and-error methods.
A recent study explores the application of deep learning in ecological resource research, addressing challenges such as multi-source/multi-meta heterogeneity and high dimensional complexity. The study highlights the potential of deep learning in connecting computer science with classical theoretical sciences in ecology.
A new AI-powered algorithm has revolutionized the analysis of deep sleep patterns by automating the detection of K-complexes. The tool, developed by Flinders University researchers, outperforms human scoring methods in speed and accuracy, providing a more comprehensive understanding of sleep health.
Researchers used AI to speed up the search for a key material in a new catalyst that converts carbon dioxide into ethylene with record efficiency. The resulting electrocatalyst has an 80% faradaic efficiency, a new record for this reaction, and shows promise for clean energy storage and carbon capture.
The human brain balances complexity and accuracy when processing patterns, with errors playing a crucial role in learning and cognition. The new model suggests that the brain constantly strives to represent things in simple terms, with participants showing quicker responses to sequences generated by modular networks.
Researchers developed an AI tool to automatically diagnose atrial fibrillation and five common ECG abnormalities, comparable to human diagnosis. The AI was trained on a large database of manually diagnosed ECGs and shows great potential for improved cardiovascular care in low-income countries.
Researchers at Carnegie Mellon University have developed a new AI-powered teaching interface that allows teachers to create intelligent tutoring systems in minutes, rather than hours. This innovation has the potential to increase the adoption of AI-based tutors and provide deeper insights into learning processes.
A team of scientists has developed a system utilizing image analysis and artificial intelligence to analyze the shape of large numbers of seeds from a single image. The trained model detected and segmented individual seeds with high accuracy and analyzed seeds of other crops, accelerating crop breeding and analysis.
Researchers successfully rebuilt bridge between experimental neuroscience and artificial intelligence learning algorithms by demonstrating a new accelerated brain-inspired learning mechanism. This mechanism outperformed commonly-used machine learning algorithms in handwritten digit recognition tasks with small datasets.
Fugu outperforms BBA in terms of least interruption time, highest image resolution and consistency of video quality, keeping viewers engaged for 5-9% longer
Researchers propose Anticipated Learning Machine (ALM) for precise future-state predictions based on short-term data. ALM efficiently reconstructs dynamics even with a small number of samples by constraining to a low-dimension space.
A team of scientists used machine learning to speed up the process of identifying optimal self-assembling peptides for biocompatible electronic devices. By screening 8,000 candidates, they were able to rank each design and pave the way for experimentalists to test the most promising ones.
Researchers used machine learning to analyze nationwide health checkup records and identified a reliable method for predicting diabetes patients. The model accurately predicted future incidence of diabetes with an overall accuracy of 94.9%.
A team of researchers used machine learning to study complex spin models, revealing key similarities between distinct phases. By training an AI on one model and applying it to another, they found that the algorithm could correctly classify phases and identify temperature transitions.
Researchers at Ruhr-University Bochum used artificial intelligence to predict the structure of thin films, reducing the need for extensive experiments. The team developed a generative model that can generate images of the surface of a layer under specific process parameters, enabling the identification of optimal material formulas.
A team of Berkeley Lab cosmologists, led by George Stein and Uros Seljak, developed a code that best identified a mock signal hidden in simulated particle-collision data. Their efficient machine learning tool, called sliced iterative optimal transport, can run on a simple desktop or laptop computer.
A new AI-driven system, DeepSPM, demonstrates fully-autonomous Scanning Probe Microscopy (SPM) operation, allowing for optimal data acquisition and quality assessment without human supervision. This breakthrough enables long-term SPM operation and bridges the gap between nanoscience, automation, and artificial intelligence.
Researchers at University of Münster develop AI tool to predict reaction outcomes using molecular structures, enabling accurate predictions for yields and stereoselectivities. The model can be applied to diverse reactions and is expected to significantly change the approach to chemical syntheses.
A new computer algorithm inspired by the mammalian olfactory system rapidly learns patterns and identifies smells even with strong sensory interference. The algorithm is applied to a neuromorphic computer chip, Loihi, which can learn to identify patterns or perform tasks a thousand times faster than traditional methods.
A new study finds that machine learning algorithms can accurately diagnose mastitis origin and reduce mastitis levels on dairy farms. The technique achieved a classification accuracy of 98% for environmental vs contagious mastitis and 78% for lactation vs dry period environmental mastitis.
Professor Gregory Ditzler is developing mathematical models and algorithms to recognize patterns and identify relevant features in machine learning. His research aims to prevent security threats in autonomous vehicles and other applications.
Rice University researchers developed a cost-saving alternative to GPU acceleration called SLIDE, which uses general-purpose CPUs without specialized hardware. The algorithm outperforms traditional back-propagation training with hash tables, reducing computational overhead and enabling faster deep learning on CPUs.
Researchers use machine learning to accelerate analysis of buried interfaces and edges in materials, creating stronger, more energy-efficient materials. The technique pairs atom probe tomography with machine learning to extract composition profiles and compare them to actual ground truth.
Researchers at USC developed personalized learning robots for children with autism, which could autonomously gauge engagement in long-term therapeutic interventions. The robots achieved 90% accuracy in detecting a child's interest in tasks.
Research finds poor connectivity between brain hubs rather than specific regions causes learning difficulties. Children with well-connected hubs have either specific cognitive difficulties or none at all, while poorly connected hubs lead to widespread and severe problems.
Professor Mary-Anne Williams from the University of Technology Sydney wins a Google TensorFlow Faculty Award to develop educational content with TensorFlow 2.0 and design machine learning experiences. The award aims to support responsible AI technologies that people can understand and trust.
A new AI algorithm developed by University of Illinois researchers accurately predicts corn yield using deep learning and convolutional neural networks. The approach incorporates various topographic variables, soil electroconductivity, nitrogen treatment rates, and seed application to optimize crop management decisions.