Tag

Machine Learning

All articles tagged with #machine learning

science-and-technology2 days ago•74 min saved

New Deep Learning Method VIMA Identifies Disease-Specific Spatial Patterns in Tissues

Researchers have introduced VIMA, a new computational method that uses deep learning to analyze spatial molecular data in tissues. By creating numerical 'fingerprints' for small tissue patches, VIMA identifies disease-associated structures without requiring manual cell segmentation or discrete clustering. Tested on rheumatoid arthritis, ulcerative colitis, and dementia datasets, the tool successfully recapitulated known biology and revealed new spatial features, outperforming existing methods in detecting case-control differences.

via Nature|

#bioinformatics #diagnostics #machine-learning

climate-and-environment8 days ago•53 min saved

Study confirms Antarctic ice loss is inevitable this century, with high emissions driving up to 25 cm sea-level rise

A new study in Nature concludes that the Antarctic Ice Sheet is committed to mass loss by 2100, even under aggressive emissions cuts. Using machine learning to analyze ice-sheet models, researchers found that higher emissions significantly increase projected sea-level rise, potentially reaching 25.4 cm by 2100 under worst-case scenarios. The study highlights that physical assumptions in models, not just emissions, drive uncertainty, and that satellite data confirms the inevitability of ice loss.

via Nature|

#antarctic-ice-sheet #climate-and-environment #climate-projections

science15 days ago•47 min saved

CERN’s CMS and ATLAS experiments tighten limits on microscopic black holes

CERN’s CMS and ATLAS experiments have found no evidence of microscopic black holes in recent proton collisions, but they have significantly narrowed the energy range where such objects could exist. Using 138 inverse femtobarns of data from 2016-2018, CMS excluded semiclassical black holes below 9.0 to 11.4 TeV, while ATLAS’s 2022-2024 data excluded models up to 9.4 TeV. These null results, achieved through new machine learning techniques, constrain theories involving extra dimensions and help resolve the hierarchy problem by eliminating viable parameter spaces for quantum gravity.

via The Brighter Side of News|

#cern #large-hadron-collider #machine-learning

technology22 days ago•5 min saved

From paper to partner: AI agents that answer questions and collaborate on research

Nature reports on Paper2Agent, a system that ingests a paper’s text, code, and data, hosts it on an MCP server, builds a paper-specific AI agent, and lets researchers interact in natural language; the agent can apply the paper’s methods to new data and even collaborate with agents from other disciplines. In tests on the AlphaGenome paper (DNA sequence predictors), the agent was created in about 45 minutes at a cost of ~$14 and achieved near-perfect accuracy on genetics questions, outperforming other biomedical AI agents and enabling reanalysis of conclusions without new experiments.

via Nature|

#ai #cross-disciplinary #machine-learning

science27 days ago•11 min saved

AI's Einstein Test: Can Vintage Models Rediscover Relativity?

Nature examines efforts to train historical-data language models to rediscover relativity, finding that current AI generally struggles to make genuine abductive leaps or paradigm shifts. While some experiments show brief hints of intuition, most results rely on data correlations and fail to produce true breakthroughs, prompting researchers to rethink model architectures and incorporate world-model reasoning, with several vintage-data projects showing limited but intriguing sparks rather than Einstein-level discoveries.

via Nature|

#artificial-intelligence #history-of-science #machine-learning

science1 month ago•50 min saved

Evolutionary trees power a genome-wide variant predictor

GPN-Star is a phylogeny-aware genomic language model trained on whole-genome alignments across vertebrates to learn functional constraints and predict genome-wide variant effects. Using a transformer with phylogeny-informed cross-attention, it achieves state-of-the-art performance in coding and non-coding variant interpretation, improves pathogenic variant prioritization and trait heritability enrichment, and extends to five model organisms. The results show that different evolutionary timescales inform different regulatory versus coding regions; while requiring WGAs for inference, GPN-Star delivers higher accuracy with modest compute compared to single-sequence models and provides public predictions and code to advance human and comparative genomics.

via Nature|

#genomics #gwas #machine-learning

science1 month ago•12 min saved

Second Fruit Fly Brain Connectome Complete, Paving Way for Neurobiology

Researchers completed the second complete connectome of a fruit fly brain: the male Drosophila connectome (≈150,000 neurons and >300 million synapses) was mapped over about four years using serial electron microscopy, AI-powered stitching and segmentation, and human proofreading in a Janelia–Google collaboration. It follows a female connectome finished earlier this year and should accelerate neurobiology research while enabling new insights into sex differences, including 289 male-specific neurons, 71 female-specific neurons, and 138 shared neurons wired differently.

via Ars Technica|

#connectome #drosophila #electron-microscopy

health1 month ago•4 min saved

AI Reveals Migraine as a Spectrum with Distinct Subtypes

Researchers used machine learning on data from 43,197 Norwegians to identify migraine signatures beyond headache symptoms, achieving strong discrimination (AUC ~0.80) without relying on traditional headache criteria. The study suggests migraine is a spectrum with four subgroups and distinct genetic signals, where age, neck pain, menstruation and nausea were key predictors; genetics offered only modest gains, hinting at personalized treatments in the future.

via Euronews.com|

#artificial-intelligence #genetics #health

science1 month ago•11 min saved

Pasadena teen uses AI to sift 200 billion infrared observations, spotting 1.5 million candidates

An 18-year-old Pasadena high school senior, Matteo Paz, built a neural-network–assisted pipeline to comb through nearly 200 billion raw infrared detections from NASA's NEOWISE survey, sorting by variability to flag about 1.5 million potential new objects. Described in The Astronomical Journal and aided by Caltech/IPAC mentorship, the method turns a data firehose into manageable leads rather than confirmed discoveries. The catalog will undergo follow-up observations to confirm which candidates are real, with Paz’s win in the 2025 Regeneron Science Talent Search marking a starting line for applying the approach to future time-domain surveys.

via Space Daily|

#astronomy #infrared-data #machine-learning

technology1 month ago•4 min saved

PlayStation 4 Blu-ray Head Sparks DIY High-Speed Dermal Microscope

Edwin Hwu demonstrates a High Speed Dermal AFM (HS-DAFM) built from a PS4 Blu-ray optical head to image skin at the nanoscale, enabling rapid dermatological and cosmetic analysis; with machine learning, the setup is explored for diagnosing conditions such as asthma (about 75% accuracy), showing a cost-effective, DIY path beyond traditional lab AFMs.

via Hackaday|

#atomic-force-microscope #dermal-imaging #machine-learning

technology1 month ago•2 min saved

AI Models Whisper to Each Other Without Words

A startup called Mostik, with input from Russian mathematicians, is testing a method for AI models to communicate with one another without human language—a 'machine telepathy' approach that could speed collaboration between models.

via WIRED|

#ai #inter-model-communication #machine-learning

science1 month ago•58 min saved

AI-Driven Seismic Scan Reveals Six Hidden Deep-Earth Patches

A study using deep learning to analyze over 2 million earthquake records (1990–2024) identified about 174,900 PKP precursor signals and mapped six previously undocumented deep-mantle regions (B1–B6) near the core–mantle boundary, suggesting large, connected pockets of heterogeneity likely tied to ancient subducted material and thermochemical piles. The findings, published in the Journal of Geophysical Research: Solid Earth, refine our view of mantle dynamics and could impact our understanding of earthquakes, volcanism, and tectonics, while calling for higher-resolution follow-up studies.

via Yahoo|

#core-mantle-boundary #deep-mantle #machine-learning

science1 month ago•7 min saved

AI Reveals Six Hidden Deep-Mantle Structures Beneath Earth

AI-assisted analysis of more than 2 million earthquake recordings (1990–2024) identified 174,929 high‑quality PKP precursor signals and unveiled six previously undocumented deep-mantle zones (B1–B6) near the core–mantle boundary. The expanded dataset shows that these small-scale structures may connect into larger belts, likely shaped by subduction and thermochemical processes, with broad implications for mantle convection, volcanic activity, and Earth's magnetic field. The researchers call for higher‑resolution, multi‑wave studies to refine deep-Earth models.

via ScienceAlert|

#core-mantle-boundary #deep-mantle #machine-learning

science1 month ago•8 min saved

Recall Shapes Memory: Retrieval Changes How We Remember

A Rice University study shows that how we retrieve information—whether focusing on a specific feature or broader category—does not drastically change recall performance, but it leaves distinct neural traces. Using fMRI and machine-learning analysis, researchers found retrieval style alters brain patterns in memory-related and perceptual areas, indicating memory is dynamic and updated during recall rather than simply replayed.

via Futurity|

#fmri #machine-learning #memory

technology1 month ago•22 min saved

WiFi as Radar: Researchers Image People and Rooms Using Unencrypted Signals

A Science Daily report says researchers used unencrypted WiFi beamforming feedback to reconstruct radio-based images of people and their surroundings, identifying individuals with nearly 100% accuracy in tests (197 participants) from multiple angles, even when the subject isn’t carrying a device. The technique relies on RF wave propagation and existing WiFi signals, meaning turning off a personal device may not stop it; once a model is trained, identification takes seconds, raising significant privacy and surveillance concerns.

via Slashdot|

#machine-learning #privacy #rf-imaging