The Knowledge Microscope: Features as Better Analytical Lenses than Neurons

2025-02-18

AI Summary

This work addresses the limitations of using individual MLP neurons as analytical units in large language models—namely, polysemy in factual knowledge representation, poor interpretability, and weak privacy protection. We propose replacing raw neurons with semantic features extracted via sparse autoencoders (SAEs) as the fundamental unit for knowledge analysis. We systematically demonstrate, for the first time, that SAE features exhibit superior unambiguity, stronger causal influence on factual recall, and enhanced capability for privacy erasure compared to neurons. Building on this, we introduce FeatureEdit—a method enabling efficient and precise knowledge editing and sensitive information removal. Experiments show that feature-level modeling significantly outperforms neuron-level approaches across knowledge representation fidelity, semantic interpretability, and privacy preservation. Notably, our approach achieves substantial gains over state-of-the-art methods in factual knowledge erasure tasks.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Previous studies primarily utilize MLP neurons as units of analysis for understanding the mechanisms of factual knowledge in Language Models (LMs); however, neurons suffer from polysemanticity, leading to limited knowledge expression and poor interpretability. In this paper, we first conduct preliminary experiments to validate that Sparse Autoencoders (SAE) can effectively decompose neurons into features, which serve as alternative analytical units. With this established, our core findings reveal three key advantages of features over neurons: (1) Features exhibit stronger influence on knowledge expression and superior interpretability. (2) Features demonstrate enhanced monosemanticity, showing distinct activation patterns between related and unrelated facts. (3) Features achieve better privacy protection than neurons, demonstrated through our proposed FeatureEdit method, which significantly outperforms existing neuron-based approaches in erasing privacy-sensitive information from LMs.Code and dataset will be available.

Problem

• Features improve knowledge expression in LMs.
• Features enhance interpretability and monosemanticity.
• Features provide superior privacy protection in LMs.

Innovation

• Sparse Autoencoders decompose neurons
• Features enhance interpretability and monosemanticity
• FeatureEdit improves privacy protection