Stable Anisotropic Regularization

TL;DR

This paper introduces I-STAR, a novel regularization method that accurately measures and adjusts isotropy in LLM embeddings, showing that decreasing isotropy can enhance model performance.

Contribution

The paper proposes I-STAR, a differentiable and stable isotropy measure, and demonstrates that reducing isotropy improves NLP model performance, challenging previous assumptions.

Findings

Decreasing isotropy improves task performance.

IsoScore* is a reliable measure of isotropy.

Regularization with I-STAR enhances training stability.

Abstract

Given the success of Large Language Models (LLMs), there has been considerable interest in studying the properties of model activations. The literature overwhelmingly agrees that LLM representations are dominated by a few "outlier dimensions" with exceedingly high variance and magnitude. Several studies in Natural Language Processing (NLP) have sought to mitigate the impact of such outlier dimensions and force LLMs to be isotropic (i.e., have uniform variance across all dimensions in embedding space). Isotropy is thought to be a desirable property for LLMs that improves model performance and more closely aligns textual representations with human intuition. However, many of the claims regarding isotropy in NLP have been based on the average cosine similarity of embeddings, which has recently been shown to be a flawed measure of isotropy. In this paper, we propose I-STAR: IsoScore*-based STable Anisotropic Regularization, a novel regularization method that can be used to increase or decrease levels of isotropy in embedding space during training. I-STAR uses IsoScore*, the first accurate measure of isotropy that is both differentiable and stable on mini-batch computations. In contrast to several previous works, we find that decreasing isotropy in contextualized embeddings improves performance on the majority of tasks and models considered in this paper.