Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers have developed Uncertainty-Aware Self-Paced Learning (UASPL), a new AI training method that uses evidential neural networks to integrate predictive reliability into sample selection. By prioritizing samples b…
IBM researchers unveiled a new generative AI model architecture called Continued Fraction Generative Networks (CoFrGeNets) at the International Conference on Machine Learning in Seoul, which fundamentally alters transfor…
A developer released a starter kit for running ONNX Runtime Web inside a Web Worker to keep browser-based machine learning inference off the UI thread, improving performance on mobile devices. The project provides reusab…
Fine-tuning adapts a pretrained large language model to a specific task using task-specific data, building on the general language understanding learned during pretraining. Pretraining teaches a model to predict the next…
Elon Litman published a technical walkthrough of Joint Embedding Predictive Architectures (JEPA), Yann LeCun's self-supervised learning method that trains models to predict representations of masked image regions from vi…
Researchers introduced NOTES, a framework combining neural operators with evolutionary strategy, achieving over 95% efficiency in nanophotonic beam-deflector design by reducing dimensionality from 256 to 25. The method o…
Researchers introduced ALER-TI, a novel time series imputation method that uses historical data patterns to reconstruct missing values more accurately than existing approaches. By aligning latent embeddings between incom…
Researchers introduced DiPhon, a diffusion framework using graphon dynamics to generate large graphs without retraining. The method trains on smaller graphs and scales to larger ones at inference, matching key statistica…
Researchers introduced DominoTree, a training-free best-first draft tree method for speculative decoding that uses Domino's conditional correction to achieve up to 6.6x speedup over autoregressive decoding on Qwen3-4B ac…
New research shows that reinforcement learning, specifically Group Relative Policy Optimization (GRPO), outperforms traditional fine-tuning for automatic speech recognition (ASR) in regulated sectors, reducing word error…
Artificial intelligence is transforming academic conferences by using deep learning and knowledge graph technology to analyze vast amounts of data, helping researchers find relevant information and predict trends. This t…
New research into deep feedforward ReLU networks reveals how path relationships and input space division mechanisms demystify the 'black box' of neural networks, offering clarity on how multiple hidden layers interact an…
A developer argues that multimodal models in production are bottlenecked by sampling decisions—frame rate, chunk length, resolution, and crop—rather than model capability. These preprocessing defaults silently determine …
AI systems cannot forget or forgive due to their engineering: weights retain training data permanently, context windows switch between perfect recall and total erasure, retrieval stores lack temporal decay, and logs pres…
Researchers developed an AI pipeline that digitizes paper ECGs via smartphone photos, achieving 95.51% accuracy in detecting myocardial infarction. The system runs on CPUs in under 30 seconds, enabling remote healthcare …
Researchers have developed quantum Wasserstein generative adversarial networks (GANs) that generate full-resolution MNIST images without dimensionality reduction or patchwork models, achieving state-of-the-art results wi…
Researchers introduced Single-rollout Asynchronous Optimization (SAO), a new reinforcement learning method for large language models that uses single-rollout sampling and double-side token-level clipping to improve stabi…
Efficient Long-hOrizon (ELO) learning, a new learned optimizer, surpasses traditional models like AdamW and Muon across diverse tasks including language modeling and image classification, requiring less than 7 H100 GPU-h…
Researchers introduced negative squaring, a pre-tilting technique that adjusts weights against expected rounding errors before quantization, achieving 77% error reduction in a toy recurrent network with 3-bit quantizatio…
Researchers introduced a bifidelity method that uses generative models to efficiently quantify uncertainties in parameter estimates, reducing computational costs without sacrificing accuracy. The approach was tested on n…