Conceptpolysemanticity
Neuron Polysemanticity
The phenomenon in which a single neuron in a neural network responds to multiple unrelated concepts simultaneously. It is an important starting point for mechanistic interpretability research and has driven the development of methods such as sparse autoencoders.
Timeline (last 90 days)
Oct 27
Individual neurons are often polysemantic, i.e. they respond to multiple unrelated concepts at the same time, so subsequent research turned to methods such as sparse autoencoders (SAEs) to extract purer feature directions.
Unverified50%