Articles tagged with Deep Learning
Hankelformer improves forecasting of extreme weather events by capturing local spatiotemporal dynamics and enhancing feature invariance. It achieves state-of-the-art performance on multiple datasets, including energy, transportation, and extreme weather domains.
Arkansas researchers used a machine-learning approach to study the organization of neighboring genes in bacteria. The novel method distinguished disease-causing strains of Enterococcus cecorum from nonpathogenic ones by analyzing how neighboring genes are organized within the bacterial genome. This new approach may provide valuable clu...
Researchers developed a deep-learning compiler to map user-designed illusion patterns to programmable metasurfaces, enabling rewritable and customizable electromagnetic illusions. The metasurface can create complex two-dimensional illusions that adapt to changing conditions.
The SNU team introduces Cluster-aware Upcycling, leveraging semantic structure of pretrained models to promote specialization among expert modules. This approach outperforms conventional Sparse Upcycling on image-text retrieval and various image classification benchmarks.
Hyunsoo Lee, an SNU undergraduate, presents research in generative visual computing at leading conferences NeurIPS, CVPR, and ECCV. His work spans image editing, human motion, and 3D content generation, leveraging pretrained generative models to produce consistent outputs.
Researchers have demonstrated the potential of artificial intelligence to make future visual prostheses, like a bionic eye, more precise and predictable. By training a deep neural network to predict patterns of brain activity produced by different electrical stimulation settings, the AI-designed patterns reproduced targeted brain activ...
Researchers developed a method for generating personality assessment questionnaires with ChatGPT from any source text, including the DSM-5 and an astrology textbook. The generated questionnaires showed high internal consistency within personality clusters and could predict life outcomes like depression, anxiety, and well-being at level...
A team of researchers at UMass Amherst has developed an AI model, DiffuDose, that generates a patient's radiation dose map with gold-standard accuracy in under 23 seconds. This technology has the potential to unlock the full potential of radiopharmaceutical therapy for prostate cancer treatment.
Researchers at Harvard John A. Paulson School of Engineering and Applied Sciences developed an AI recommendation model that incorporates reinforcement learning to adjust to the uniqueness of each user. This approach improved human-AI performance over traditional one-size-fits-all decision support.
A new USC study uses AI to generate detailed maps of brain aging, revealing distinct patterns of neurodegeneration in specific regions. The approach sheds light on how local brain age correlates with changes in cognitive function across the lifespan.
Researchers developed a novel AI framework that optimizes investment decisions directly while accounting for risk. The study found that conventional forecasting-based approaches were outperformed by the decision-focused model in terms of risk-adjusted performance and wealth accumulation.
A study by Bar-Ilan University researchers found that learning is driven primarily by changes in the strength of existing neural connections. The models became significantly better at learning as the amount of training data increased, but the proportion of lost connections remained roughly the same.
The review explores how integrating Federated Learning (FL), Reinforcement Learning (RL), and Natural Language Processing (NLP) can overcome modern NLP system limitations, such as protecting user privacy and adapting to changing environments. The study presents a unified framework that combines FL, RL, and NLP as three co-equal pillars.
A new AI model called Biogeochemistry-Informed Neural Network (BINN) has been developed to advance scientific discovery in agriculture and biogeochemistry. It is 50 times more efficient than its predecessors and can predict biological processes not yet well understood.
A study proposes an operational framework combining machine learning models with Sentinel-2 data to estimate agricultural drought conditions in irrigated and non-irrigated maize fields. Deep Neural Network (DNN) achieved the best performance, showing higher prediction accuracy and lower error metrics for non-irrigated fields.
Researchers created a virtual robot with curiosity-driven neural network and tested its performance, finding that play-like behavior and exception-handling performance helped the robot understand language faster. The study suggests a combination of curiosity and linguistic diversity is key to children's rapid language acquisition.
Researchers developed an AI platform, PeptiVerse, to predict key properties of peptides, enabling early assessment of drug potential. The open-source platform allows users to evaluate ordinary and chemically modified peptides, streamlining the discovery process.
Researchers developed a high-resolution monitoring approach combining drone-based multispectral imaging with ensemble machine learning models. The method revealed how surrounding industrial, agricultural, residential, and green areas influence water quality in urban rivers.
A novel AI model called BINND has been developed to predict which DNA molecules bind to each other. The model achieved an accuracy of 83.5% in predicting DNA pairs that would bind, surpassing the state-of-the-art model by at least 10%. This improvement has significant utility for biomedical diagnostic tools and DNA computing applications.
Deep learning models accelerate drug design, predict chemical interactions, and engineer stable candidates. AI-powered simulations optimize dosimetry, predicting biodistribution and generating patient-specific digital twins for individualized treatment planning.
Researchers at the University of Pennsylvania and Chinese University of Hong Kong created TD3B, an AI framework guiding peptide generation toward candidates predicted to have a desired effect. The tool predicts binding likelihood and determines activation or deactivation of associated cellular machinery.
Researchers have developed a federated learning algorithm that solves the long-standing conflict between robustness and efficiency in AI development. The new approach anonymizes data and reduces single-point failure risks while maintaining speed. By remembering past client interactions, servers can protect against malicious input.
Researchers developed an AI approach that identifies nanoparticle morphology using data from standard NTA measurements, achieving high classification accuracies. The method integrates two types of information and performs multi-class classification with stable performance even at reduced data amounts.
A new learning-based adaptive tuning method integrates chaotic search with particle swarm optimization to improve stability and solution quality in chaotic search algorithms. The approach consistently achieves better results than conventional methods, providing a practical means of enhancing the performance of chaotic search.
Researchers at Michigan State University found that current AI models can be duped into seeing signatures of life in digital organisms with high accuracy. However, when tested on unseen examples, the results were less impressive. The team showed that it was possible to convince the AI that it was seeing signs of life where they didn’t ...
A large-scale study of an online patient portal shows that AI-generated responses can introduce errors and extraneous details, leading to increased editing time for physicians. Adapting AI to individual physician communication styles can improve accuracy by 33% and reduce editing by 26%.
Researchers uncover a previously unknown phase transformation mechanism in monolayer molybdenum telluride (MoTe2) that is fundamentally distinct from the conventional martensitic model. The study reveals a one-dimensional 'domino-like' chain reaction that triggers structural rearrangement and enables programmable electronic devices.
Researchers have delivered a comprehensive roadmap for deep learning-based face video restoration, offering a unified framework to understand and advance the rapidly evolving field. Dedicated face video restoration techniques can restore clarity, preserve identity, and maintain temporal smoothness across video frames.
Researchers developed a compact, cost-effective diagnostic platform combining lensfree holography and deep learning for automated HER2 scoring. The system reached 84.9% accuracy for four-level HER2 classification and 94.8% accuracy for binary scoring, effectively lowering diagnostic risks.
Researchers used AI to analyze mammograms and found that women who developed breast cancer had increasing risk scores over time, while those who did not had stable scores. The study suggests that image-based AI risk scores can predict future breast cancer risk in women without a known genetic mutation or family history.
A University of Houston engineering professor developed a mathematical model to help decision-makers decide where to spend limited dollars on infrastructure resilience. The model accounts for real-world uncertainty and identifies critical assets to invest in, providing the greatest benefit before disaster strikes.
A Chinese research team has developed a deep learning model that can predict the South Indian Ocean Dipole (SIOD) seven months in advance, outperforming traditional dynamical forecasting systems. The model uses sea surface temperature and ocean heat content anomalies as inputs and automatically learns key features of ocean temperature ...
A groundbreaking technology called Time to Move (TTM) offers unprecedented control over object and character movement in AI-generated videos. TTM eliminates the need for complex infrastructure or training on millions of videos, making AI video creation more accessible.
A Concordia-led team developed an AI-based method for detecting toxic online content, which outperformed existing tools in accuracy and throughput. The Proximal Policy Optimization-based Cascaded Inference System (PPO-CIS) layers scanning tasks to quickly identify harmful material.
Avishek Choudhury, a WVU researcher, has won the NSF CAREER award to study how healthcare providers' trust in artificial intelligence changes over time. His goal is to humanize algorithms behind AI and improve decision-making quality and patient safety.
A novel operator learning framework tackles the ill-posedness in inverse thermal problems by employing two synergistic deep learning components. The framework accurately estimates spatial material properties and filters high-frequency noise to project the final-state thermal field back to its initial distribution.
FireANTs, an open-source algorithm, combines AI optimization and geometry to quickly match complex medical images. The new method can accomplish what took weeks in minutes, detecting subtle changes that signal disease or cognitive decline, making it practical for clinical practice.
TurboLynx, developed by POSTECH researchers, analyzes complex, interconnected data up to 184 times faster than existing systems. The engine groups similar data together and processes them collectively, reducing unnecessary memory usage and enabling efficient analytical queries.
Researchers used deep learning to model energy release during r-process nucleosynthesis in hydrodynamic simulations, gaining new insights into element formation. The results suggest that r-process heating is an important effect that should be better accounted for in future modeling.
A new study found that AI-powered chatbots can make vaccine-hesitant parents more likely to say they will immunize their children against HPV, but no more than standard written public health materials. Additionally, the effects of the chatbots did not last longer than those of government health materials.
A new study published in the Journal of Big Data highlights the journal's emergence as a leading publication in data science and artificial intelligence research. The study found that JBD has become a central hub for high-impact research worldwide, with significant contributions from top researchers.
Binghamton University researchers have developed a new way to reduce troublesome fake information in AI chatbots, with high accuracy in identifying disease terms and drug names. The protocol harnesses multiple large language models to verify answers through 'voting', increasing confidence in the results.
A deep learning model combines knowledge from different catalyst families to identify a top-performing green hydrogen catalyst. The AI correctly predicted the activity ranking of 12 tested catalysts within a previously unexplored material family.
A new framework, SUVA, enables organizations to measure and adjust AI chatbots' social preferences, improving their performance in customer complaints and other human-AI interactions. By understanding an LLM's existing tendencies, organizations can decide whether an available model already fits its values and usage scenarios.
A new AI system, Empirical Research Assistance (ERA), can automatically write scientific software programs that outperform human-written ones. ERA combines a large language model with search strategies to explore and refine thousands of pieces of code, reducing the time required for exploration from months to hours or days.
A new model combines text mining and machine learning to extract service-specific aspects and customer actions from online reviews. The model effectively identifies core technical issues and user love for a platform, enabling targeted decisions for improvement. Researchers validated the model using 231,705 online reviews of Roblox.
Researchers have developed a framework called CHEEM that allows AI models to learn new tasks without losing performance on existing tasks. The framework improves adaptive intelligence by tailoring computational structure depending on task complexity.
A team of researchers proposes a deep learning architecture called CCDNN to learn correlated representations for multi-source data fusion. The method demonstrates promising performance, surpassing existing methods in reconstruction tasks and achieving better results in industrial fault diagnosis and remaining useful life cases.
A research team led by POSTECH developed an AI framework that can predict and account for microscopic defects in metal 3D printing, improving the reliability of metal components. The framework achieves a Mean Absolute Error (MAE) of just 9.51 MPa, outperforming conventional approaches.
Recent advances in noninvasive neural decoding, deep learning, and shared autonomy are bringing BCI-controlled robots closer to real-world use. These systems can decode complex brain signals related to movement intention, allowing for more flexible and higher-dimensional robotic behaviors.
Researchers developed ApexGO, an AI-powered method to turn weak antibiotic candidates into more potent ones. The tool uses generative AI and Bayesian optimization to guide molecular tweaks, predicting which changes are likely to increase antimicrobial activity.
Researchers developed a deep learning method that optimizes both material layout and fiber direction for continuous fiber composite structures, reducing design time by up to 99.7% while maintaining strong accuracy. The ResUNet-GAN framework produces high-performance structural layouts directly from design parameters.
Micro-gesture recognition is emerging as a key area of research in affective computing, focused on analyzing subtle, involuntary body movements that may reflect hidden emotional states. The field has expanded from RGB video to skeleton, audio, text, and privacy-preserving multimodal settings.
Researchers have developed a deep-learning framework to reconstruct a global, high-resolution nighttime light dataset from 1992 to 2024. The new product improves upon existing datasets by reducing saturation-related bias and better capturing temporal changes in urbanization, economic activity, and human development.
Researchers developed a machine learning approach to analyze Fermi surface images, identifying compositions with significant changes and nodal lines. The method accurately detects outliers, enabling efficient screening of large datasets for desirable electronic properties.
Researchers have developed SmartDJ, an AI-powered editor that allows users to reshape audio experiences with simple words. The system uses language models and diffusion models to interpret high-level requests and generate edited outputs.
Researchers developed a cancer assessment tool that can identify high-risk patients and specific cell populations linked to their risk. The tool, called scSurvival, predicts survival outcomes more accurately than traditional methods by analyzing single-cell data at cellular resolution.
Researchers propose a novel SENet-CNN-Transformer model to predict electric vehicle charging duration, outperforming existing models in accuracy and reducing training time. The approach combines data enhancement, channel attention, convolutional neural networks, Transformer modeling, and transfer learning to address real-world data sca...
A review article highlights a deep learning-driven CNN approach for detecting and classifying dynamic road obstacles, achieving high accuracy in obstacle identification and classification. The proposed architecture shows strong performance, but real-world deployment requires continued evaluation across larger and more varied scenarios.
Researchers developed a health-aware V2G strategy using reinforcement learning to optimize charging and discharging times, resulting in significant lifecycle cost savings ($1,539) and extended battery life (21 months). The study suggests electric bus charging stations can be promising platforms for scalable V2G services.