Word
polysemanticity
Definition
n. the property of a single artificial neuron firing in response to several completely unrelated concepts.
n. the phenomenon whereby an individual neuron or activation dimension responds selectively to multiple distinct, semantically unrelated concepts or visual features.
Examples
“A single neuron showed polysemanticity by activating for pictures of cats, guitars, and German text.”
“The existence of polysemanticity makes it very difficult for humans to understand what individual neurons are calculating.”
“Sparse autoencoders address polysemanticity by decomposing dense entangled activations into a much larger overcomplete dictionary of interpretable monosemantic directions.”
Examples
simple
“A single neuron showed polysemanticity by activating for pictures of cats, guitars, and German text.”
contextual
“The existence of polysemanticity makes it very difficult for humans to understand what individual neurons are calculating.”
complex
“Sparse autoencoders address polysemanticity by decomposing dense entangled activations into a much larger overcomplete dictionary of interpretable monosemantic directions.”
Real-World Examples
“Polysemanticity is a phenomenon in neural networks in which individual neurons respond to multiple unrelated concepts rather than to a single well-defined one.” “This paper provides a toy model where polysemanticity can be fully understood, arising as a result of models storing additional sparse features in "superposition.” Etymology
Coined in the 2020s by mechanistic interpretability researchers at Anthropic and OpenAI from Greek poly- ('many') + semantikos ('meaningful'), by analogy with linguistic polysemy.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/polysemanticity