Polysemanticity
The phenomenon in which a single neuron in a neural network responds to multiple unrelated features. It is a core obstacle in the field of mechanistic interpretability, making it difficult for researchers to clearly map internal network activations to specific concepts.
Timeline (last 90 days)
The same neuron may participate in multiple different functions, the so-called "polysemanticity," so intervening in it produces unexpected side effects.
Early interpretability research found that a certain neuron activated strongly to both cat faces and car fronts, and this phenomenon was called polysemanticity.
All Facts (2)
The same neuron may participate in multiple different functions, the so-called "polysemanticity," so intervening in it produces unexpected side effects.
50%UnverifiedEarly interpretability research found that a certain neuron activated strongly to both cat faces and car fronts, and this phenomenon was called polysemanticity.
50%