Word
monosemanticity
Definition
n. the ideal property where an individual feature or neuron corresponds to only one clear, understandable concept.
n. the property of a neural representation where each latent feature direction corresponds unambiguously to a single isolated and human-interpretable concept.
Examples
“Scientists strive for monosemanticity so each feature in the model can be explained with a single label.”
“Sparse dictionary learning successfully extracted millions of features exhibiting high monosemanticity from the frontier model.”
“Achieving true monosemanticity across all hidden layers is considered the holy grail of mechanistic interpretability and AI model auditing.”
Examples
simple
“Scientists strive for monosemanticity so each feature in the model can be explained with a single label.”
contextual
“Sparse dictionary learning successfully extracted millions of features exhibiting high monosemanticity from the frontier model.”
complex
“Achieving true monosemanticity across all hidden layers is considered the holy grail of mechanistic interpretability and AI model auditing.”
Etymology
Coined in the 2020s by Anthropic interpretability researchers as the direct antonym to polysemanticity, from Greek mono- ('single') + semantikos ('meaningful').
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/monosemanticity