Part of Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Main Conference Track
Ryan Singh, Christopher L Buckley
Attention mechanisms play a crucial role in cognitive systems by allowing them to flexibly allocate cognitive resources. Transformers, in particular, have become a dominant architecture in machine learning, with attention as their central innovation. However, the underlying intuition and formalism of attention in Transformers is based on ideas of keys and queries in database management systems. In this work, we pursue a structural inference perspective, building upon, and bringing together, previous theoretical descriptions of attention such as; Gaussian Mixture Models, alignment mechanisms and Hopfield Networks. Specifically, we demonstrate that attention can be viewed as inference over an implicitly defined set of possible adjacency structures in a graphical model, revealing the generality of such a mechanism. This perspective unifies different attentional architectures in machine learning and suggests potential modifications and generalizations of attention. Here we investigate two and demonstrate their behaviour on explanatory toy problems: (a) extending the value function to incorporate more nodes of a graphical model yielding a mechanism with a bias toward attending multiple tokens; (b) introducing a geometric prior (with conjugate hyper-prior) over the adjacency structures producing a mechanism which dynamically scales the context window depending on input. Moreover, by describing a link between structural inference and precision-regulation in Predictive Coding Networks, we discuss how this framework can bridge the gap between attentional mechanisms in machine learning and Bayesian conceptions of attention in Neuroscience. We hope by providing a new lens on attention architectures our work can guide the development of new and improved attentional mechanisms.