Neural Networks

Researchers have developed AGFlow, a spatiotemporal flow-matching model that fuses asynchronous Sentinel-1 radar and Sentinel-2 optical satellite data to remove clouds and reconstruct missing time-series frames. The mode…

Researchers have developed STROP, a discrete visual tokenizer that learns to generate variable-length structural scene representations, determining the optimal sequence length for each image in a single forward pass. The…

Researchers have introduced Tensor Memory, a lightweight module that gives Transformer models a fixed-size recurrent 3D memory tensor to handle long video sequences. The module writes tokens into a voxel grid using a dif…

A study evaluating FP4 quantization-aware training (QAT) for real-time anomaly segmentation in brain tumor detection found that architecture choice, not QAT recipe, most determines model quality. The Swin Transformer dem…

Researchers have developed representation-conditioned diffusion models that generate synthetic training images by conditioning on learned representations from DINOv2, DINOv3, and CLIP, achieving a 10.76 percentage point …

Researchers at the University of Cambridge have developed the first large-scale Ukrainian handwritten text generation system, creating a dataset of 126,177 images from 308 writers and retraining a diffusion-based model t…

Researchers have developed a new interpretation method for Transformer models that use heterogenous attention structures, which process information from different sources. The approach addresses challenges in understandi…

NVIDIA's CUTLASS library, a header-only C++ template framework for writing custom CUDA kernels, powers much of the AI infrastructure behind FlashAttention, vLLM, and PyTorch's internal kernels. The library enables develo…

DTensor, PyTorch's distributed tensor abstraction, attaches placement metadata to every tensor to automatically propagate layouts and insert correct collective operations during distributed training. While this approach …

Researchers from Sakana AI and the University of Tokyo developed DiffusionBlocks, a block-wise training framework that converts residual networks into independently trainable denoising modules. The method partitions a ne…

Working memory capacity could be expanded by applying radio signal processing techniques to neural activity, allowing the brain to encode and combine abstract concepts through frequency modulation. Researchers propose th…

A developer built babelForge, a neuroscience tool designed to map, model, simulate, and project factors affecting neurobiological systems. The project originated from a research initiative for PTSD, created after the dev…

Sreeni Ramadorai draws a parallel between Transformer models' residual connections and Agile software development sprints, arguing both systems preserve foundational knowledge while incrementally adding improvements. The…

A multidisciplinary team led by Professor Douglas Kelley at the University of Rochester used physics-informed neural networks to infer brain-wide glymphatic fluid velocities from MRI tracer videos. The method, called MR-…

A developer explained that tensors are the fundamental data structures used in AI systems to represent information, ranging from 0-dimensional scalars for single values to n-dimensional tensors for complex data like vide…

A biologically plausible stochastic gradient descent (SGD) algorithm is required to simulate the effect of additional brain mass for human intelligence augmentation, but the correct local learning rule remains unidentifi…

A developer discovered that adding skip connections—the line `x = x + output`—solved the problem of deeper neural networks producing more errors, a challenge documented in a Microsoft Research paper. This technique, whic…

PyTorch's Inductor compiler uses kernel fusion to accelerate model execution by up to 10x, grouping dependent operations into single Triton kernels to reduce memory traffic and kernel launch overhead. In a pointwise fusi…

A developer conducted controlled experiments on CIFAR-10 using a Convolutional Neural Network and found that data preprocessing alone caused model accuracy to range from 65% to over 87%, with one case dropping to nearly …

Researchers have identified brain-computer interfaces (BCIs) as a viable path to create moderately superintelligent humans, capable of outperforming current civilization's best minds in alignment and x-risk cognition by …