Neural Networks

Researchers introduced FAST, a framework that accelerates Temporal Graph Neural Network training by integrating sampling, memory I/O, and computation optimizations, achieving up to 4.7x speedup without accuracy loss. The…

ManifoldFlow, a new approach to neural network weight constraints, introduces flexible spectral control by allowing a bounded positive spectrum instead of fixed singular values, outperforming traditional Stiefel layers i…

Researchers introduced OrthoReg, a method that penalizes overlap between physics-based equations and neural networks in hybrid models, improving symbolic recovery and out-of-distribution performance in dynamical systems.…

Researchers have reimagined Transformers as a dynamic dance between fast and slow paths, where weight dynamics and stability play essential roles. The framework divides parameter space into visible and invisible directio…

Researchers propose a 'shell-core' theory to explain grokking, a delayed generalization phenomenon in neural networks, where solutions move from an outer shell to a core of generalization. The framework suggests that usi…

Researchers developed a hybrid quantum-classical framework combining LSTMs with Quantum Circuit Born Machines to improve financial volatility forecasting. The model outperformed classical counterparts on high-frequency d…

A new cross-attention Transformer model enhances Wi-Fi signal decoding by jointly processing uplink OFDM signals without explicit channel estimates, outperforming traditional methods and neural baselines in realistic tes…

A study on deep neural network security testing examined four coverage metrics across LeNet, VGG, and ResNet architectures, finding that deeper models and larger datasets do not automatically improve security coverage. T…

Researchers developed GenShin, an AI model that predicts protein corona composition on lipid nanoparticles without costly mass spectrometry, potentially accelerating drug delivery innovation. The geometry-enhanced graph …

Researchers developed MambaCapsule, an AI model that combines Mamba and Capsule networks to diagnose cardiac arrhythmia with 99.54% accuracy on the MIT-BIH dataset and 99.59% on the PTB dataset, while providing transpare…

Researchers introduced the input-skip LBBNN (ISLaB) technique, which reduces neural network complexity by over 99% while maintaining accuracy, achieving 97% accuracy on MNIST with only 935 weights. The method enhances in…

Neural networks and logistic regression models are competing to improve stroke prediction, as strokes occur every 40 seconds in the U.S. and are a leading cause of death. Neural networks offer complex pattern recognition…

Order-equivariant neural networks (OENN) are advancing AI by leveraging symmetry to improve deep learning efficiency and performance. Researchers have characterized linear order-equivariant maps, constructed OENN layers,…

New research shows that 12-31% of weights in sparse transformers can be clearly interpreted, compared to dense models. An automated pipeline generates human-readable descriptions for each weight, testing them on unseen d…

Researchers have introduced Threshold Gating, a new activation method for neural networks that unifies traditional functions like ReLU and Sigmoid under a single framework. The approach allows pre-trained networks to be …

A researcher trained a character-level transformer on Nietzsche texts and identified a copying head (B5H0) using eigenvalue analysis, but found its QK circuit insufficient for bracket completion. The study disconfirmed h…

A new benchmarking framework for Graph Explainable AI (G-XAI) evaluates Graph Neural Networks (GNNs) using tabular explainability metrics that separate graph topology and node features, identifying non-dominated explaine…

A developer reports issues with their LSTM implementation in C++, noting that training error initializes around 4.5 but can spike into the 20s, which they have not seen with other networks. They suspect a matrix indexing…

New research challenges assumptions about infinite-width neural networks, introducing a width-solid learnability theorem that shows infinite width does not guarantee an edge over polynomial width for learning specific ta…

Researchers have developed physics-informed neural networks (PINNs) that embed the axisymmetric equations of linear elasticity to model wave propagation in bimaterial systems like steel and aluminum. The approach accurat…