Stochastic Parroting in Temporal Attention -- Regulating the Diagonal Sink

TL;DR

This paper investigates the bias in temporal attention mechanisms caused by over-squashing, introduces sensitivity bounds, and proposes regularization methods to mitigate the diagonal sink effect in spatio-temporal models.

Contribution

The paper provides a theoretical analysis of diagonal attention sink in temporal attention and introduces regularization techniques to address it.

Findings

Diagonal attention matrices suffer from a sink effect.

Regularization methods improve temporal attention performance.

Sensitivity bounds relate off-diagonal scores to sequence length.

Abstract

Spatio-temporal models analyze spatial structures and temporal dynamics, which makes them prone to information degeneration among space and time. Prior literature has demonstrated that over-squashing in causal attention or temporal convolutions creates a bias on the first tokens. To analyze whether such a bias is present in temporal attention mechanisms, we derive sensitivity bounds on the expected value of the Jacobian of a temporal attention layer. We theoretically show how off-diagonal attention scores depend on the sequence length, and that temporal attention matrices suffer a diagonal attention sink. We suggest regularization methods, and experimentally demonstrate their effectiveness.