nGPT: Normalized Transformer with Representation Learning on the Hypersphere Article Swipe

PDF

Related Concepts

Hypersphere Transformer Representation (politics) Mathematics Computer science Artificial intelligence Engineering Electrical engineering Political science Voltage Politics Law

Ilya Loshchilov , Cheng-Ping Hsieh , Simeng Sun , Boris Ginsburg ·

YOU? · · 2024 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2410.01131 · OA: W4403883328

We propose a novel neural network architecture, the normalized Transformer (nGPT) with representation learning on the hypersphere. In nGPT, all vectors forming the embeddings, MLP, attention matrices and hidden states are unit norm normalized. The input stream of tokens travels on the surface of a hypersphere, with each layer contributing a displacement towards the target output predictions. These displacements are defined by the MLP and attention blocks, whose vector components also reside on the same hypersphere. Experiments show that nGPT learns much faster, reducing the number of training steps required to achieve the same accuracy by a factor of 4 to 20, depending on the sequence length.