Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new study reveals that existing class unlearning methods in machine learning ignore class geometry, enabling Class Membership Inference Attacks (CMIA) that leak information about supposedly forgotten data. Researchers …
A new study on robust AI planning introduces non-asymptotic convergence bounds for Q-learning and actor-critic algorithms in distributionally strong Markov Decision Processes, achieving an optimal strong policy with accu…
An experiment transforms 1D data into 3D topology for applications including AI audio deepfake detection, protein folding, drug docking, solar magnetic flares, and music visualization.
A developer outlines a practical workflow for transcribing audio to text, covering automated, manual, and hybrid methods. The workflow includes preprocessing audio with ffmpeg, running OpenAI Whisper locally, and exporti…
Google's TPU clusters use 2D or 3D torus topologies with 4 or 6 nearest neighbors per chip, while NVIDIA GPU clusters rely on fat-tree topologies over InfiniBand, according to a July 14, 2026 technical deep dive. The pos…
Researchers have discovered that deep neural models for click-through rate predictions suffer a sharp decline in validation performance after just one training epoch, even as training loss improves, due to overfitting. B…
A new neural network optimization method using a symmetric-exponential weight reparameterization achieved up to 30% faster training on transformers, reaching equivalent validation loss in 1.32 to 1.49 times fewer steps o…
A new feedback-driven approach enhances Retrieval Augmented Generation (RAG) systems by integrating human input to iteratively improve response quality, according to a recent study. The method adds an auxiliary feedback …
A new regulation-driven fine-grained hierarchical classification framework has been proposed to improve accuracy and transparency in compliance tasks such as customs tariff classification and export control categorizatio…
CLIR-Bench, a new benchmark built from de-identified ICU records with 6,600 QA instances across 11 clinical variables, reveals that existing generalist AI models struggle to reason over irregular clinical time series, hi…
A major Eastern European bank has deployed a transformer-based foundation model that unifies multimodal user event sequences into a single chronological stream, outperforming traditional predictive models on financial ta…
A developer created a derivative dataset from the Orthogonal Model of Emotions by removing one dimension and replacing ontological cluster labels with 10 emoji, then validated it using Create ML with Transfer Learning ov…
Researchers have found that the multimodal language model LLaVA 1.6 can replicate six of seven categories of naming errors seen in stroke-induced aphasia, matching real-world clinical data in 97.8% of configurations, tho…
A study on on-device English-to-Traditional-Chinese subtitle translation shows that reducing the vocabulary of the LMT-60-0.6B model from 151,000 to 64,000 tokens improves speed and privacy. The LocalSubs model outperfor…
Adaptive Model Compression (AMC) reduces energy consumption by 59.2% and boosts throughput 2.24x on 45nm CMOS edge devices, with only a 3.6% accuracy trade-off. The saliency-driven framework dynamically allocates hardwar…
The SCOPE-RL framework, a two-stage reinforcement learning approach for large language models, improved average accuracy by up to 11.2 percentage points and reduced reasoning tokens by 27.1% on models like Qwen3-8B-Instr…
A senior software engineer and tech lead, drawing on frameworks from Dan Shapiro and Nate B. Jones, argues that AI skills built for human use degrade under agent load in predictable ways, producing output that looks corr…
A new study analyzing data from 27,375 UK Biobank participants using MRI scans and machine learning found that metabolic syndrome significantly accelerates brain ageing, with the effect increasing with the number of meta…
An open-source pipeline called ai-trains-ai uses nested reinforcement learning loops where an outer agent is RL-trained to write inner training jobs for small models, with the entire system run for roughly $1.3k on real …
A conjecture known as the 'pseudpocalypse' warns that anyone who writes significant amounts of text on the internet can have their pseudonymous identities linked through statistical analysis of their writing style, with …