Dong, YiheCordonnier, Jean-BaptisteLoukas, Andreas2021-09-252021-09-252021-09-252021-01-01https://infoscience.epfl.ch/handle/20.500.14299/181645WOS:000683104602074Attention-based architectures have become ubiquitous in machine learning. Yet, our understanding of the reasons for their effectiveness remains limited. This work proposes a new way to understand self-attention networks: we show that their output can be decomposed into a sum of smaller terms-or paths-each involving the operation of a sequence of attention heads across layers. Using this path decomposition, we prove that self-attention possesses a strong inductive bias towards "token uniformity". Specifically, without skip connections or multi-layer perceptrons (MLPs), the output converges doubly exponentially to a rank-1 matrix. On the other hand, skip connections and MLPs stop the output from degeneration. Our experiments verify the convergence results on standard transformer architectures.text::conference output::conference proceedings::conference paper