Cluster-norm for Unsupervised Probing of Knowledge

TL;DR

This paper introduces a cluster normalization method to improve unsupervised probing of language models by reducing the influence of unrelated features, thereby enhancing the accuracy of knowledge extraction.

Contribution

The paper proposes a novel cluster normalization technique that enhances unsupervised probing accuracy by mitigating the effects of distracting features in language models.

Findings

Improved accuracy of unsupervised probes in identifying intended knowledge

Reduction of misleading features in probing results

Enhanced robustness of knowledge extraction methods

Abstract

The deployment of language models brings challenges in generating reliable information, especially when these models are fine-tuned using human preferences. To extract encoded knowledge without (potentially) biased human labels, unsupervised probing techniques like Contrast-Consistent Search (CCS) have been developed (Burns et al., 2022). However, salient but unrelated features in a given dataset can mislead these probes (Farquhar et al., 2023). Addressing this, we propose a cluster normalization method to minimize the impact of such features by clustering and normalizing activations of contrast pairs before applying unsupervised probing techniques. While this approach does not address the issue of differentiating between knowledge in general and simulated knowledge - a major issue in the literature of latent knowledge elicitation (Christiano et al., 2021) - it significantly improves the ability of unsupervised probes to identify the intended knowledge amidst distractions.