ML//neural network//polysemanticity

A single neuron responds to multiple unrelated concepts: the same neuron fires for "cat photos", "the concept of trust", and "words ending in -tion". This is the default state of neural networks, not the exception.

A single neuron responds to multiple unrelated concepts: the same neuron fires for "cat photos", "the concept of trust", and "words ending in -tion". This is the default state of neural networks, not the exception.

The direct consequence of superposition: when a model encodes more concepts than it has neurons, concepts must share neurons. Each neuron participates in multiple unrelated representations, creating polysemantic activation patterns.

This is why naive neuron-level interpretability fails. Looking at what maximally activates a single neuron gives you a confusing collage of unrelated inputs. The neuron isn't "about" any single thing; it's a shared axis in a compressed space.

Monosemanticity is the goal, polysemanticity is the obstacle. The work of mechanistic interpretability is largely about decomposing polysemantic neurons into monosemantic features using tools like sparse autoencoders

The "grandmother cell" debate in neuroscience is the same question: does the brain have one neuron per concept (monosemantic) or does each neuron participate in many concepts (polysemantic)? In artificial networks, the answer is clear: polysemantic is the norm because it's more parameter-efficient.

Polysemanticity scales with model capacity: smaller models have MORE polysemantic neurons (more compression), larger models trend toward more monosemantic representations (more capacity). This partially explains why larger models are easier to interpret.