Unveiling and Controlling Anomalous Attention Distribution in Transformers

TL;DR

This paper investigates the high attention on the first element in Transformer models, analyzing its causes and proposing methods to control this anomalous attention distribution for improved model understanding and efficiency.

Contribution

It introduces a novel analysis of the waiver phenomenon causing attention anomalies, categorizing waiver element selection based on positional encoding and feature distribution.

Findings

High attention on first element is linked to waiver phenomenon.

Two methods of waiver element selection identified: positional-encoding-based and feature-distribution-based.

Analysis aids in developing techniques for attention distribution control.

Abstract

With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the development of techniques focusing on attention distribution, such as Key-Value (KV) Cache compression and infinite extrapolation; however, the latent cause leaves to be unknown. In this paper, we analyze such a phenomenon from the perspective of waiver phenomenon, which involves reducing the internal values of certain elements in the sequence, allowing them to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods: positional-encoding-based and feature-distribution-within-elements-based.