Loading…
Glossary term
Interpretability work that tries to reverse-engineer the specific circuits inside a neural network.
Mechanistic interpretability tries to reverse-engineer a neural network the way an engineer would an unknown machine: finding the features it represents and the circuits that combine them to perform a task.
Techniques include probing activations and training sparse autoencoders to find interpretable features.
It aims at a precise, causal understanding of models, which could help detect hidden problems.