Superposition Hypothesis
A theoretical hypothesis explaining polysemanticity in neural network neurons. It proposes that when the number of neurons is insufficient to represent all concepts, the model folds multiple sparsely activated features into the same neuron, resulting in representational superposition.
Timeline (last 90 days)
In 2020, Cammarata and Chris Olah provided evidence in the paper "Zoom In" showing that the network hides car feature signals inside a group of dog neurons, reinforcing the superposition hypothesis.
Neel Nanda points out that there are two kinds of superposition, arbitrary superposition and representational superposition, and what Cammarata refuted is arbitrary superposition, while representational superposition remains the core controversy.
All Facts (2)
In 2020, Cammarata and Chris Olah provided evidence in the paper "Zoom In" showing that the network hides car feature signals inside a group of dog neurons, reinforcing the superposition hypothesis.
50%UnverifiedNeel Nanda points out that there are two kinds of superposition, arbitrary superposition and representational superposition, and what Cammarata refuted is arbitrary superposition, while representational superposition remains the core controversy.
50%