Mechanistic interpretability

Mechanistic interpretability (often abbreviated as mech interp, mechinterp, or MI) is a subfield of research within explainable artificial intelligence that aims to understand the internal workings of neural networks by analyzing the mechanisms present in their computations. The approach seeks to analyze neural networks in a manner similar to how binary computer programs can be reverse-engineered to understand their functions.

History

The term mechanistic interpretability was coined by Chris Olah.^[1]

Early work combined various techniques such as feature visualization, dimensionality reduction, and attribution with human-computer interaction methods to analyze models like the vision model Inception v1.^[2]

Key concepts

Mechanistic interpretability aims to identify structures, circuits or algorithms encoded in the weights of machine learning models.^[3] This contrasts with earlier interpretability methods that focused primarily on input-output explanations.^[4]

Linear representation hypothesis

Simple word embeddings exhibit a linear representation of semantics. The relationship between a country and its capital is encoded in a linear direction in the example above.

This hypothesis suggests that high-level concepts are represented as linear directions in the activation space of neural networks. Empirical evidence from word embeddings and more recent studies supports this view, although it does not hold up universally.^[5][6]

Methods

Mechanistic interpretability employs causal methods to understand how internal model components influence outputs, often using formal tools from causality theory.^[7]

Mechanistic interpretability, in the field of AI safety, is used to to understand and verify the behavior of complex AI systems, and to attempt to identify potential risks.^[8]