Neural Networks

A new study improving the Huth et al. fMRI language decoding pipeline achieved a mean METEOR score of 0.149 and BLEU-1 score of 0.200, an 11% relative METEOR gain over baseline, but the fMRIFlamingo model's 42.86% Top-1 …

A new investigation into transformer models for reinforcement learning reveals that larger embedding dimensions produce more sophisticated internal world models and improve interpretability, even though small embeddings …

Researchers have released Deep4ge, a massive dataset containing 14,227 training runs from 59 deep neural network programs to help detect faults in deep learning systems. The dataset includes 9,845 faulty runs created usi…

A new framework, VMoGE (Variational Mixture of Graph Neural Experts), promises to enhance the accuracy of dementia diagnosis using multi-band EEG analysis, achieving an AUC of 0.89 in distinguishing healthy controls from…

Researchers have developed a physics-informed neural network (PINN) that predicts tension during parachute line extraction, outperforming traditional numerical methods in both computational efficiency and precision. The …

A new approach using sparse autoencoders (SAEs) is revolutionizing out-of-distribution (OOD) detection by extracting interpretable features from intermediate neural network layers, achieving record performance on OOD det…

A new training-free attribution method reveals that Feedforward Networks in Transformer models exhibit sparse, structured inter-layer dependencies, where only a small subset of preceding neuron activations significantly …

New research reveals that up to 77% of features in degraded Sparse Autoencoders (SAEs) are functionally inert, and even in well-trained models 9% of matched features are dead, challenging the interpretability claims of S…

A new AI-driven beamforming framework using an enhanced in-context learning Transformer backbone with encoder-decoder networks outperforms existing methods in multi-user MISO systems without retraining across different c…

A study on arXiv (2607.11967v1) found that donor-specific functional fingerprints remain detectable in neural networks after Neural Collapse, with all 20 ordered donor-recipient pairs correctly identified (p=0.0083) usin…

Researchers have developed a self-evolving in-context learning (ICL) framework for pilot-to-beamformer design in multi-user multiple-input single-output (MU-MISO) systems that outperforms existing methods including the W…

A study comparing semantic memory navigation in blind and sighted individuals found that sighted participants showed higher entropy for abstract concepts, while blind participants did not, instead exhibiting higher entro…

Researchers at arXiv improved the Huth et al. fMRI encoding baseline by expanding voxel selection to 15K and substituting GPT-2 medium for GPT-1, achieving mean METEOR = 0.149 and BLEU-1 = 0.200 for subject UTS03, an 11%…

A new arXiv preprint (2607.11958v1) from researchers studying the free energy principle demonstrates that a predictive system's default behavior can decouple from its representational accuracy, a phenomenon called ontolo…

A study comparing the RG-Flow Transformer, which incorporates a renormalization-group inductive bias, against a parameter-matched vanilla transformer on scarce EEG data from the PhysioNet Sleep-EDF corpus found no signif…

Researchers introduced Semidirect Fourier Delta Attention (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control, achieving exact affine …

Researchers introduced a multi-feature fusion framework for non-invasive semantic reconstruction from brain recordings, achieving state-of-the-art performance by combining static lexical representations (W2V) with dynami…

OmniPM-Net, a Convolutional Conditional Neural Process-based fusion model, reconciles discrete station forecasts with gridded fields for PM10 forecasting, matching the accuracy of graph neural networks (mean absolute err…

Researchers propose SpikeDS, a spiking neural network that uses dual sparsity to predict perineural invasion from 3D MRI in cholangiocarcinoma patients. In a clinical cohort of 139 patients, SpikeDS achieves an AUC of 0.…

Object detection models are deep learning architectures that classify and localize target objects within images or video frames, outputting bounding box coordinates, class labels, and confidence scores. Two-stage detecto…