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 University of Helsinki develop a new privacy-aware machine learning method that enables accurate modeling using private user device data. This method ensures limited information on each data subject is revealed, making it ideal for protecting sensitive health and human behavior data.
The Conversational Intelligence Challenge Finals saw six teams compete, with two winning teams developing modules for processing different scenarios and a command module that decides which routine should step in. The competition aimed to work out approaches for chatbot evaluation and human-to-machine dialogue collection.
A new collaborative engineering project funded by NSF aims to make numerical computation of departure rates and route choice faster, enabling rapid rerouting and diversion. The project uses machine learning to develop statistical models of traffic flow, potentially reducing congestion by seconds, minutes or hours ahead of time.
Youth football players undergo brain changes, including alterations in the default mode network, after a single season of play. This study suggests that repeated subconcussive impacts can have a lasting effect on brain health.
Researchers at UCLA developed a deep-learning-based technique to reconstruct holograms for microscopic images, producing better results than current methods. This approach could aid in diagnosing abnormalities in medical images and improve optical microscopy for medical diagnostics.
Researchers at Numenta propose a new theory for how the brain learns models of objects through movement, pairing sensory input with location signals. The theory predicts that even the first levels of processing in the brain are learning and recognizing complete objects.
CANDLE, a scalable deep learning framework developed by Argonne's Exascale Deep Learning and Simulation Enabled Precision Medicine for Cancer project, was recognized with the award. The framework has achieved impressive results, including explaining over 92% of variance in drug response.
Researchers efficiently used Stampede2's 1024 Skylake processors to complete a 100-epoch ImageNet training with AlexNet in 11 minutes, setting the fastest time recorded to date. The Layer-Wise Adaptive Rate Scaling (LARS) algorithm enabled this breakthrough, allowing for larger-than-ever batch sizes and adaptive learning rate adjustments.
A study published in Biodiversity Data Journal uses deep learning techniques to differentiate between similar plant families with high accuracy. The researchers trained neural networks on digitized herbarium specimens, achieving up to 99% accuracy in distinguishing between two challenging species.
Researchers from Lehigh University and Columbia University have developed a new testing approach for deep learning platforms used in self-driving cars, malware-detection, and other systems. Their method, called DeepXplore, exposes thousands of unique incorrect corner-case behaviors, enabling faster identification and fixing of errors.
A new approach brings transparency to self-driving cars and other self-taught systems by automatically error-checking neural networks. Researchers found thousands of bugs missed by previous techniques, activating up to 100% of network neurons and improving accuracy up to 99%.
Researchers found that mothers alter their voice timbre when speaking to infants, regardless of native language. This consistent shift enables computers to distinguish between infant- and adult-directed speech.
Scientists developed a machine-learning method to predict molecular behavior, which can aid in developing new pharmaceuticals and enhancing emerging battery technologies. The method combines physics, chemistry, and machine learning, allowing it to simulate complex chemical behavior within molecules.
Researchers develop a quantum perovskite material that exhibits adaptive response to repeated proton insertion and removal, resembling brain's desensitization. This property enables effective programming of the material like a computer.
Carnegie Mellon University's Center for Human Rights Science has received a $100,000 grant from Open Society Foundations to explore the positive impact of emerging technologies on human rights. The grant will enable the center to catalog and share developing technologies with the broader human rights community.
Researchers have developed a web app capable of producing 3D facial reconstruction from a single 2D image. The technique, using Convolutional Neural Networks, allows for arbitrary facial poses and expressions, with over 400,000 users already trying it out.
Researchers have developed a new method to simulate infrared spectra using artificial neural networks, reducing simulation time from thousands of years to minutes. This breakthrough enables the analysis of complex chemical systems and paves the way for widespread adoption in various fields.
The study uses image data to reconstruct the cell cycle of white blood cells and the progress of diabetic retinopathy, demonstrating the method's capability in handling continuous biological processes. The software also identifies individual categories and assigns measured data to clusters when data is not part of a continuous process.
A computer science approach using machine learning predicts the time remaining before a fault fails by analyzing acoustic signals emitted during laboratory-created earthquakes. The technique identifies new signals, previously thought to be low-amplitude noise, that provide forecasting information throughout the earthquake cycle.
The Lehigh project aims to build bridges between optimization experts, learning theorists, and statisticians to advance machine learning. With a $1.5 million grant, the team will develop innovative educational pathways and state-of-the-art mathematical tools for data science, promoting long-term research and training activities in the ...
A team of ORNL researchers aims to use deep learning to identify patterns in scientific data that alert scientists to potential new discoveries. They plan to leverage ORNL's Titan supercomputer and develop novel high-performance computing methods.
Researchers at Disney Research developed an algorithm that can automatically recognize soccer formations and defensive strategies from player tracking data. The algorithm outperformed conventional methods in identifying dynamic player roles and coordinated team behavior, with applications beyond sports.
Researchers have developed a new algorithm that allows AI to collect error reports and correct them immediately, without affecting existing skills. This enables robots to learn from their mistakes and spread new knowledge amongst themselves.
Kolachalama's research aims to improve cardiovascular treatments by developing models for smarter artery care and improving drug-coated angioplasty balloons. The $231,000 Scientist Development grant will support his three-year study on the mechanisms of drug-coated balloon therapy.
A new AI system developed by MIT researchers reduces online video rebuffering and improves quality, adapting to different network conditions. The system achieves higher-quality streaming with less rebuffering than existing approaches.
A new study proposes that cultural activities, such as language use, affect our ability to collect data, make connections, and infer behavior. The research reveals that the brain's limited working memory can be beneficial in some cognitive tasks, unlike our closest relatives, chimpanzees.
A team of researchers from the University of Texas at Austin has developed novel approaches to information retrieval that leverage artificial intelligence, crowdsourcing, and supercomputing. Their method combines input from multiple annotators to determine the best overall annotation for a given text, improving accuracy in extracting d...
Researchers used active machine learning to discover new conditions for synthesizing gigantic polyoxometalate molecules. The algorithm outperformed human experimenters, covering a broader range of the 'crystallization space' and discovering unexpected crystals.
Researchers at the University of British Columbia have developed a breakthrough algorithm called DeepLoco that enables computer characters to learn complex motor skills like walking and running through trial and error. The system uses deep reinforcement learning to allow characters to respond to their environment without hand-coding st...
Researchers found that babies as young as 7 months can learn a second language with just one hour of play-based exposure per day. The study used an English-language method and curriculum in Madrid's public infant-education centers, showing significant improvements in English comprehension and production.
Researchers developed a new algorithm that can turn audio clips into highly-realistic videos of people speaking, using available public domain video footage. This technology has potential applications in improving video conferencing and creating realistic virtual reality experiences.
The University of Texas at San Antonio is developing an artificial neural network called NFrame to monitor and detect 'bad behavior' in computer systems. The system will learn normal behaviors and flag anomalies, allowing it to predict potential issues and prevent security breaches.
Researchers aim to develop a system that can explain AI decisions to humans, making autonomous vehicles and robots more trustworthy. The project will utilize real-time strategy games like StarCraft to train AI players that can provide natural language explanations.
Researchers at Tsinghua University outline recent advances on nonparametric Bayesian methods, regularized Bayesian inference, scalable algorithms, and system implementation to tackle the challenges of Big Data. They also discuss connections with deep learning and highlight the need for human expertise in devising appropriate features a...
Machine learning techniques are successfully applied to image-based diagnosis, disease detection, and prognosis in medical imaging. The review focuses on denoising methods using machine learning approaches to develop a systematic decision for diagnosing and prediction of medical images.
The Biogerontology Research Foundation is helping to develop artificial intelligence for accelerated drug discovery in aging and age-related diseases. Researchers have made significant progress in using deep learning-based approaches to characterize biomarkers of ageing and predict the chronological age of patients.
Researchers have developed a method that enables computers to infer psychologically plausible models of individuals from limited data, improving understanding of human behavior. This breakthrough could lead to more accurate predictions and adaptations in human-robot interaction and adaptive interfaces.
Researchers trained AI models to identify TB on chest X-rays, achieving a 96% accuracy rate. The models' performance was improved when combined with expert radiologist interpretation, increasing the diagnosis accuracy to nearly 99%.
Researchers found that machine learning programs can acquire cultural biases from online language patterns, affecting tasks like image categorization and automated translations. This study highlights the importance of identifying and addressing bias in AI systems to promote fairness and equality.
Researchers at Stanford University have created a deep learning algorithm that can accurately predict the toxicity of chemicals and associate drugs with side effects using just six data points. This breakthrough could help chemists choose promising candidates and accelerate drug development.
Researchers at UTA developed OnTask, a software tool that offers timely and personalized feedback to help students adjust their studying throughout the course. The online tool uses data about student activities and suggests strategies to increase confidence and success.
Researchers have developed a machine learning model that can predict the outcome of cellular interactions and design new cancer treatments. The Stampede supercomputer enabled the team to run billions of simulations, allowing them to identify patterns in the data and create a system capable of predicting laboratory results.
Researchers used deep learning and machine learning techniques to analyze S&P 500 data, achieving statistically significant and economically substantial returns. The findings challenge the efficient-market hypothesis and suggest AI can excel in times of high volatility.
A study of Google DeepMind's access to millions of NHS patient records reveals concerns over data privacy and regulation. The researchers argue that the original agreement lacked transparency and suffered from an inadequate legal basis, serving as a cautionary tale for public sector bodies and private tech firms.
A machine learning application powered by artificial intelligence has been developed to provide evidence-based answers to frequently asked questions in interventional radiology. The system enables real-time communication between clinicians and patients, improving the quality of care.
A new automated method based on deep learning techniques analyzes detailed game data to create models of how a typical player would behave in a given situation. This allows for the comparison of actual player behavior with predicted ghostly behavior, providing valuable insights into defensive athletic performance.
Researchers have developed software to decode digital brain data, enabling rapid progress on the ability to monitor neural activity and understand learning, memory, and cognitive functions. The collaboration has reduced processing time from days to less than a second.
A new UT Dallas study derived optimal policies and data-driven techniques for firms to learn about demand and adjust capacity. The study's main finding is that production managers need to maintain a careful balance between observing the demand and changing the capacity.
Swiss researchers use neural networks to challenge the resolution limit of telescopes, recovering features that were previously invisible. The technique, inspired by a generative adversarial network, achieves better results than previous methods, such as deconvolution, and has vast potential for future astronomical observations.
A Northwestern University and Los Alamos National Laboratory team developed a novel workflow to design new materials with useful electronic properties. By combining machine learning and density functional theory calculations, they created design guidelines for ferroelectricity and piezoelectricity.
A new project by Upside Energy and Heriot-Watt University aims to use machine learning and distributed AI to manage storage assets and provide real-time energy reserves. The goal is to relieve stress on the grid and reduce reliance on traditional power stations.
Researchers have identified nearly two-dozen solid electrolytes that could replace volatile liquids in smartphones and laptops. The AI-powered approach allows for rapid screening of materials, identifying the most promising candidates for further study.
Researchers are preparing to tackle an onslaught of petabytes of complex data from sophisticated experiments, including CERN's Large Hadron Collider. To keep up with the challenge, experts propose developing exascale supercomputers and smarter networks, as well as reengineering software to adapt to future hardware developments.
A team of researchers from the University of Toronto has developed a novel machine learning approach to determine whether planetary systems are stable or not. This method is 1,000 times faster than traditional methods and can provide valuable information about exoplanets, including their mass and orbital eccentricity.
Researchers at U of T Engineering developed an AI algorithm that learns directly from human instructions, exceeding conventional training methods by 160% and outperforming its own training by 9%. The algorithm's potential lies in applying heuristic training to fields like medicine and transportation.
Researchers developed a machine learning classifier to discover membrane-active peptides with diverse sequences. The approach identified new peptides with broad biomedical implications, including immunotherapy and anticancer therapeutics.
Researchers at Numenta compared their biologically-derived HTM sequence memory to traditional machine learning algorithms, demonstrating comparable prediction accuracy. The new paper highlights the algorithm's properties, including continuous online learning and robustness to sensor noise, making it ideal for streaming data applications.
Researchers developed a new AI system that uses web search to extract data from plain text, outperforming traditional machine learning methods by up to 10%. The system learns to generate search queries and gauge relevance, then extracts relevant data from online texts.
Researchers at UTA are developing an AI system that assesses children's executive function skills, recognizing patterns of inattention and hyperactivity. The system provides recommendations for effective intervention and monitoring progress over time.
A new data-cleaning tool called ActiveClean analyzes a user's prediction model to identify mistakes and update the model as it works. By minimizing human error, ActiveClean improves model accuracy and reduces statistical biases, making it an essential tool for building better prediction models.