linear-representation-hypothesis-definition
IN premise — summaries/2026/08/24/elhage-2022-toy-models-superposition-chunk-3.md
Created 2026-08-25T02:57:58+00:00
The Linear Representation Hypothesis states that neural network features are encoded as specific directions in activation space, with the map from features to activation vectors being linear even though the features themselves are nonlinear functions of the input.
Summary
Even though a neural network applies complex nonlinear transformations to its inputs, the internal features it builds line up as simple directional "pointers" in its activation space, meaning each feature corresponds to a specific axis you can measure along. This matters because it gives us a practical handle: we can probe, manipulate, or remove a feature with a single linear step, even inside a network whose overall computation is anything but linear.