The concept that a many unrelated features are packed within the same few neurons.

  • The neuron that fires for the ear of a dog might be the same that fires for the door of a car Neurons do not describe one specific feature. They are “re-used” for different purposes. Polysemanticity encodes information that are unlikely to be mutually present in the same input.

Cases

  • InceptionV1 - cat faces and cars encoded in same neuron

Solutions