Tag: #neural-architecture5 results
ITNet is a novel learnable integral transform that mathematically subsumes convolutional networks, transformers, and recurrent models. By using a learnable kernel implemented as an MLP, it enables a single architecture to adapt its behavior to various data modalities.
Princeton researchers discovered that the brain uses modular 'Cognitive Lego' neural patterns to combine existing knowledge for new tasks, offering a potential blueprint for solving AI's catastrophic forgetting.
Novel architecture uses damped harmonic oscillators for token processing, handling local context via resonance. Achieves 1.34 BPB on FineWeb at 14.8M params, stable training, and quantization robustness. Code in 300 lines of PyTorch; seeking arXiv endorsement.
Researchers have published a proof demonstrating that EML-type trees can approximate any function within a reasonably general space. The study utilizes a construction of elementary functions as building blocks to achieve universal approximation capabilities.
Autochess NN is a browser-playable neural chess engine reaching ~2700 Elo, built on a home RTX 4090 using AI-assisted research. It features residual CNN + transformer architecture, trained on 100M+ positions via Lichess pretraining, Syzygy fine-tuning, and self-play RL. Includes play-vs-AI, puzzles, and move analysis with CPU inference under 2ms.