#deep learning

'Compare Focal Loss and Binary Cross-Entropy on a 99:1 imbalanced dataset to see how Focal Loss improves minority-class detection and yields more meaningful decision boundaries.'

Sora, OpenAI's video generator, is under scrutiny after outputs resembled copyrighted Netflix and TikTok content, sparking legal and ethical debates about scraped training data.

'A practical comparison of PyTorch and TensorFlow in 2025 covering developer experience, performance, deployment ecosystems, and use case guidance to help you choose the right framework.'

Google AI and UC Santa Cruz Genomics Institute released DeepPolisher, a deep learning tool that substantially reduces errors in genome assemblies, improving the accuracy of human genome references.

A detailed technical comparison of Alibaba's Qwen3 30B-A3B and OpenAI's GPT-OSS 20B MoE transformer models, highlighting architectural differences and use case recommendations.

MIT researchers have developed a method to stabilize large transformer training by enforcing Lipschitz bounds through spectral weight regulation and the Muon optimizer, eliminating the need for traditional normalization techniques.

Falcon-H1 from TII introduces a hybrid model combining attention and state space mechanisms, achieving performance on par with leading 70B parameter LLMs while optimizing efficiency and scalability.

GenSeg is a novel generative AI framework that significantly enhances medical image segmentation performance in scenarios with limited labeled data by creating optimized synthetic datasets.

Google DeepMind’s new AI, Aeneas, assists historians by analyzing ancient Latin inscriptions, offering dating, origin insights, and text restoration suggestions to enhance epigraphic research.

Radial Attention introduces a novel sparse attention mechanism that cuts training costs by 4.4× and inference time by 3.7× in video diffusion models, enabling generation of longer videos without quality loss.

Google DeepMind's AlphaGenome is a novel deep learning model that predicts the regulatory impact of DNA mutations across multiple biological modalities with high precision, outperforming existing models in genomic tasks.

Discover the top artificial intelligence books recommended for 2025, covering foundational concepts, advanced techniques, ethical issues, and future trends in AI.

A novel AI framework introduces differentiable MCMC layers that enable neural networks to efficiently learn with inexact combinatorial solvers, significantly improving performance in complex optimization problems like vehicle routing.

Meta introduces KernelLLM, an 8-billion-parameter model that automates converting PyTorch modules into efficient Triton GPU kernels, outperforming larger models in kernel generation benchmarks.

Mila & Universite de Montreal researchers introduce FoX, a novel Transformer variant with learnable forget gates that improve long-context language modeling efficiency and accuracy without computational trade-offs.

Microsoft researchers demonstrate that the Muon optimizer drastically speeds up grokking in Transformer models, enabling faster transition from memorization to generalization compared to AdamW.