Neural Networks

Researchers developed a controllability-observability framework to reduce hidden-state redundancy in deep neural networks, achieving up to 72.95% state compression on MNIST with minimal accuracy loss and 70.94% compressi…

New smooth Monte-Carlo estimators can evaluate the Euler characteristic in continuous neural fields 250 times faster than traditional persistent-homology methods, but they require strict design rules and sacrifice accura…

Researchers developed a Tri-Branch Modular Fusion Neural Network that integrates 3D spatial geometry, discrete topological grammar, and macroscopic physicochemical descriptors to predict molecular properties. On the QM9 …

A study of neural-interface models reveals a critical gap between formal robustness certificates and real-world safety, with EEGNet's classification accuracy dropping 25.7% under attack while certificates remained valid.…

Researchers studying neural networks on modular arithmetic found that embedding compression lags behind generalization by at least 10,000 steps, challenging traditional convergence metrics. The grokking transition overst…

Scientists developed a label-efficient AI model using Siamese neural networks that predicts phase transitions in physics with minimal data, achieving accuracy on par with established methods without requiring critical re…

Researchers have developed a new numerical method that combines neural networks with Wiener chaos decomposition and the Euler scheme to solve backward stochastic differential equations (BSDEs), offering potential improve…

Researchers have developed a new method for neural network depth adaptation that uses error estimation from continuous-time optimal control problems to dynamically insert layers where errors peak. The approach outperform…

Researchers introduced the Riemann-normal-coordinate Levenberg-Marquardt (RNC-LM) method for nonlinear optimization, which improves convergence and robustness by incorporating geodesic acceleration to handle parameter-ef…

Researchers have developed a spectral preprocessing method using FFT that significantly improves attention mechanisms in character-level language models, achieving up to a 79% reduction in validation loss on the TinyShak…

Researchers have introduced Bayesian deep ensembles, a new method that combines interpretability with high performance in predictive modeling. The approach uses low-dimensional ensemble representations and closed-form Ba…

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…

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…

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…

A developer built a Double-Tower Gated Transformer Network (GTN) to generate profitable trading signals from high-frequency EUR/USD forex data. The system uses a custom cost matrix and weighted cross-entropy loss to comb…

Researchers developed STST-JEPA, a self-supervised transformer model that predicts brain age from EEG data with a mean absolute error of 3.06 years, outperforming previous methods. The model, pretrained on over 47,000 se…

University of Tokyo researchers identified a neural mechanism in mice that links social memory to fear, causing aversion to specific individuals. The study found strengthened connections between the hippocampus and amygd…

A developer published a visual and practical guide to large language models, explaining that they are computer programs trained on massive text datasets to predict the next word. The guide covers how models like GPT-4 wo…

Researchers applied a Spectral Temporal Graph Neural Network (StemGNN) to predict user equipment scheduling states in 5G networks, recovering 57-73% of sum rate loss caused by backhaul latency. The model achieved 87.57% …

Researchers developed a Q-learning-based adaptive retraining method to address traffic drift in Open Radio Access Networks (O-RAN), reducing retraining overhead while maintaining forecasting accuracy. The approach, which…