All articles tagged with #machine learning
Billions are pouring into AI-enabled scientific research, but early experiments—such as AI models used to select mRNA at Lila Sciences—produce puzzling results, highlighting ongoing debates about whether AI can truly conduct scientific discovery.
UC San Diego researchers used two weeks of smartwatch data and EMA mood logs to train a personalized machine‑learning model that identifies each participant’s top mood drivers and pairs them with tailored, remote coaching to create an individualized Mood Augmentation Plan (iMAP). Over six weeks, 55% of participants showed depression remission on PHQ-9, anxiety decreased 36%, and benefits persisted for three months post-intervention, suggesting a scalable, data‑driven approach to personalized depression care.
A new AI suite called MouseMapper uses foundation-model–based 3D imaging to automatically segment nerves, immune cells, and 31 organs across an entire mouse body, enabling multi-system analyses of obesity. Built on VesselFM, it comprises a Nerve-Module, Immune-Module, and Tissue-Module that generalize across imaging resolutions and labeling strategies. In diet-induced obesity, the framework reveals reduced whole-body nerve density, increased nerve presence in adipose tissue, and notable infraorbital nerve remodeling in the trigeminal nerve linked to sensory deficits. Spatial proteomics of the trigeminal ganglion shows actin cytoskeleton and inflammation pathway changes, conserved in obese humans. Additionally, MouseMapper generates whole-body inflammation maps by profiling Cd68+ immune-cell clusters across tissues, illustrating systemic, organ-wide perturbations and offering a scalable bridge from cell-level changes to whole-body disease phenotypes.
A large study of about 500,000 UK Biobank participants using organ-specific aging clocks and machine-learning analysis found a U-shaped association between sleep duration and biological aging: both short (<6 hours) and long (>8 hours) sleep correlated with faster aging across organs such as the brain, heart, lungs, and immune system, while 6.4–7.8 hours per night was associated with healthier aging patterns. The work links sleep patterns to mental health and a range of diseases but does not establish causality; it suggests sleep duration is a modifiable factor in a coordinated brain–body aging process.
New cosmological analyses using supernovae, galaxy surveys, and machine-learning reconstructions reveal small but persistent deviations from the standard FLRW description of a uniform, isotropic universe. If confirmed, these effects—potentially driven by Dyer-Roeder light propagation and cosmic backreaction from large-scale structures—could challenge the Lambda-CDM framework and require new physics or revisions of how space-time evolves, with future DESI, Euclid, and other surveys poised to test the results.
Scientists are inching toward weather‑style forecasts for volcanic eruptions, but predicting eruptions with that level of certainty remains challenging because magma sits deep and each volcano is unique. Advances in seismology, ground deformation monitoring, gas measurements, and machine learning are enabling earlier warnings and more detailed volcano models. Projects like Ex-X and SZ4D seek to uncover the governing physics, improve data collection, and develop archetype volcano models that could one day output probabilistic eruption forecasts days or weeks in advance, but achieving a generalized, reliable forecast will require decades of data and a far more extensive global monitoring network.
ARC Raiders' anti-cheat program combines kernel-level protection with ML-driven telemetry (via a partnership with Anybrain) to detect cheats, while refining detection of legitimate accessibility device use by analyzing intent, and maintaining a human-reviewed ban appeals process with ongoing updates.
A new Human Organ Atlas using hierarchical phase-contrast tomography (HiP-CT) at the European Synchrotron Radiation Facility delivers detailed 3D images of 87 organs from 54 donors, exposing cellular-level anatomy and disease features (including aspects of COVID-19 and cancer) with unprecedented precision. The dataset era—exceeding terabytes—aims to support medical training, research, and AI model development, with the broader goal of eventually imaging entire bodies at 10–20x higher resolution than today’s capabilities, potentially transforming anatomy study and diagnosis.
A new arXiv study uses a machine‑learning algorithm to sift through 83,717,159 stars observed by NASA’s TESS, uncovering 11,554 exoplanet candidates (10,052 of which are newly identified) with orbital periods from 0.5 to 27 days. Researchers even confirmed a hot Jupiter, TIC 183374187 b, with the Magellan telescope, validating the method. If these candidates are verified by independent surveys, the total number of known exoplanets could rise to about 18,000, nearly triple the current count. Most candidates lie around very faint stars and require extensive follow‑up; the work posted on arXiv on April 20 has not yet been peer‑reviewed. While many candidates are unlikely to host life due to their close orbits, this study dramatically expands the census of exoplanets and demonstrates the power of machine‑learning in astronomy.
A 2022 Mauna Loa eruption, studied with a mix of government and commercial satellites and machine learning, yielded 3D lava-flow models and cooling timelines. Researchers say this Earth-based analogue helps interpret Venus data to identify ongoing volcanic activity, informing future missions like VERITAS.
MIT CSAIL researchers propose RLCR (Reinforcement Learning with Calibration Rewards), using the Brier Score to penalize over-confident wrong answers and reward well-calibrated uncertainty, leading to more reliable LLM outputs and the ability to admit when they don’t know—potentially increasing safety for high-stakes applications.
Gary Marcus argues that ChatGPT’s new image engine is visually impressive but does not demonstrate true understanding. He points to labeling errors in bike diagrams and odd results from a custom tandem-bike prompt as evidence that the system can imitate understanding without grasping how parts function. The piece emphasizes that regurgitating images isn’t the same as real comprehension in AI.
Cybersecurity experts warn that Anthropic’s Claude can generate code with vulnerabilities, underscoring the risk of relying on AI for production software without thorough review. The piece emphasizes the need for rigorous auditing and secure coding practices when AI assists with coding to prevent exploitable flaws in real-world applications.
Startups are turning videos of people doing chores (filmed by gig workers who can earn up to $25/hour) into training data for robot control software, using footage of laundry folding, dishwashing, and more to teach robots how to interpret sensor input and decide movements. The approach blends human videos, teleoperation, and simulated data to scale robot learning, a process that experts say is data-intensive and costly, with real-world deployment still years away.
Researchers found that 124 papers used two Kaggle datasets to train stroke- and diabetes-prediction models that may be built on fabricated data; some models are already in clinical use in Indonesia, Spain, and the US, with journals investigating; irregular data patterns—such as unreal completeness and duplicated values—cast doubt on reliability, prompting calls for data-source disclosure and removal of the dubious datasets to prevent flawed clinical decisions.