Derek Larson

Overview One of the active research areas for interpretability involves distilling neural network activations into clean, labeled features. This is made difficult because of superposition, where a neuron may fire in response to multiple, disparate signals making that neuron polysemantic. To date, research has focused on one type of such...