Softmax Attention with Constant Cost per Token

arXiv:2404.05843v1 Announce Type: new
Abstract: We propose a simple modification to the conventional attention mechanism applied by Transformers: Instead of quantifying pairwise query-key similarity with scaled dot-products, we quantify it with the logarithms of scaled dot-products of exponentials. Attention becomes expressible as a composition of log-sums of exponentials that is linearizable, with a latent space of constant size, enabling sequential application with constant time and space complexity per token. We implement our modification, verify that it works in practice, and …

abstract arxiv attention cost cs.cl cs.lg key per products query simple softmax token transformers type