a polysemantic neuron
/ pol-ee-suh-MAN-tik /
It would be lovely if each neuron in a model meant one tidy thing — this one fires for cats, that one for the colour red. Early interpretability hoped for exactly that. Instead, when you actually inspect a neuron, you often find it lights up for a baffling grab-bag: cat faces, the start of a legal document, and the Japanese language, all at once. A neuron that responds to several unrelated concepts like this is called polysemantic — 'many-meaning'.
Why would a network build such confusing parts? The leading explanation is superposition. Because the model wants to track more features than it has neurons, it cannot afford to dedicate a whole neuron to each concept. So a single neuron ends up participating in many different feature-directions, firing whenever any of the concepts it shares duty for is present. The opposite, rare ideal — one neuron, one concept — is called a monosemantic neuron, and most neurons in real models are nowhere near that clean.
Polysemanticity is one of the central obstacles to reading a model. If neurons each mean several things, you cannot interpret the network just by labelling neurons one by one, and a feature you care about for safety may not live in any single neuron at all. This is precisely the problem dictionary learning and sparse autoencoders are designed to attack: rather than interpreting neurons, they try to recover the underlying monosemantic features hidden inside the polysemantic mess.
Inspect one neuron in a vision model and you might see it activate for cat faces, the fronts of cars, and cat legs. It is not 'the cat neuron'; it is one switch shared across several feature codes — a textbook polysemantic neuron.
Polysemantic: one neuron, several unrelated meanings — usually a symptom of superposition.
Do not read a neuron's label as the model's true concept. The monosemantic feature you want is usually a direction across many neurons, not the neuron itself.