Disentangling Neuron Representations with Concept Vectors
This addresses a key bottleneck in mechanistic interpretability for AI researchers, though it is incremental as it builds on existing work with concept vectors.
The paper tackles the challenge of interpreting polysemantic neurons in neural networks by developing a method to disentangle them into concept vectors that represent distinct features, with evaluations showing these vectors encode coherent, human-understandable features.
Mechanistic interpretability aims to understand how models store representations by breaking down neural networks into interpretable units. However, the occurrence of polysemantic neurons, or neurons that respond to multiple unrelated features, makes interpreting individual neurons challenging. This has led to the search for meaningful vectors, known as concept vectors, in activation space instead of individual neurons. The main contribution of this paper is a method to disentangle polysemantic neurons into concept vectors encapsulating distinct features. Our method can search for fine-grained concepts according to the user's desired level of concept separation. The analysis shows that polysemantic neurons can be disentangled into directions consisting of linear combinations of neurons. Our evaluations show that the concept vectors found encode coherent, human-understandable features.