InceptionV1
A classic convolutional neural network architecture proposed by Google, widely used for image classification tasks. It has also become an important subject of analysis in mechanistic interpretability research.
Timeline (last 90 days)
Neuron 4e:55 in the InceptionV1 model is strongly activated by both cat faces and car fronts, a phenomenon long regarded as evidence that neural networks are not interpretable.
Nick Cammarata's analysis points out that InceptionV1's 4e:55 neuron actually detects a V-shaped geometric form, rather than cat faces or car fronts.
Cammarata's extreme stress experiment showed that after compressing 480 curve-detecting neurons in a certain layer of InceptionV1 down to 10 and retraining, these neurons were still clearly interpretable rather than chaotically superposed.
All Facts (3)
Neuron 4e:55 in the InceptionV1 model is strongly activated by both cat faces and car fronts, a phenomenon long regarded as evidence that neural networks are not interpretable.
50%UnverifiedNick Cammarata's analysis points out that InceptionV1's 4e:55 neuron actually detects a V-shaped geometric form, rather than cat faces or car fronts.
50%UnverifiedCammarata's extreme stress experiment showed that after compressing 480 curve-detecting neurons in a certain layer of InceptionV1 down to 10 and retraining, these neurons were still clearly interpretable rather than chaotically superposed.
50%