A copy of this work was available on the public web and has been preserved in the Wayback Machine. The capture dates from 2022; you can also visit <a rel="external noopener" href="https://arxiv.org/pdf/2110.09419v2.pdf">the original URL</a>. The file type is <code>application/pdf</code>.
<span class="release-stage" >pre-print</span>
Multi-head, key-value attention is the backbone of the widely successful Transformer model and its variants. This attention mechanism uses multiple parallel key-value attention blocks (called heads), each performing two fundamental computations: (1) search - selection of a relevant entity from a set via query-key interactions, and (2) retrieval - extraction of relevant features from the selected entity via a value matrix. Importantly, standard attention heads learn a rigid mapping between<span class="external-identifiers"> <a target="_blank" rel="external noopener" href="https://arxiv.org/abs/2110.09419v2">arXiv:2110.09419v2</a> <a target="_blank" rel="external noopener" href="https://fatcat.wiki/release/zkwt2m4nt5h6tfz3zrcrpyqwd4">fatcat:zkwt2m4nt5h6tfz3zrcrpyqwd4</a> </span>
more »... and retrieval. In this work, we first highlight how this static nature of the pairing can potentially: (a) lead to learning of redundant parameters in certain tasks, and (b) hinder generalization. To alleviate this problem, we propose a novel attention mechanism, called Compositional Attention, that replaces the standard head structure. The proposed mechanism disentangles search and retrieval and composes them in a dynamic, flexible and context-dependent manner through an additional soft competition stage between the query-key combination and value pairing. Through a series of numerical experiments, we show that it outperforms standard multi-head attention on a variety of tasks, including some out-of-distribution settings. Through our qualitative analysis, we demonstrate that Compositional Attention leads to dynamic specialization based on the type of retrieval needed. Our proposed mechanism generalizes multi-head attention, allows independent scaling of search and retrieval, and can easily be implemented in lieu of standard attention heads in any network architecture.
<a target="_blank" rel="noopener" href="https://web.archive.org/web/20220218040617/https://arxiv.org/pdf/2110.09419v2.pdf" title="fulltext PDF download" data-goatcounter-click="serp-fulltext" data-goatcounter-title="serp-fulltext"> <button class="ui simple right pointing dropdown compact black labeled icon button serp-button"> <i class="icon ia-icon"></i> Web Archive [PDF] <div class="menu fulltext-thumbnail"> <img src="https://blobs.fatcat.wiki/thumbnail/pdf/b8/b8/b8b813111c411ae61881ab9cd25707d9de6444ec.180px.jpg" alt="fulltext thumbnail" loading="lazy"> </div> </button> </a> <a target="_blank" rel="external noopener" href="https://arxiv.org/abs/2110.09419v2" title="arxiv.org access"> <button class="ui compact blue labeled icon button serp-button"> <i class="file alternate outline icon"></i> arxiv.org </button> </a>