2 ms·
Just finished reading the paper. TLDR; a linear classifier is trained on the activation values of an already trained network when it is fed examples of a speci
by goldenkey 8y ago
Just finished reading the paper.
TLDR; a linear classifier is trained on the activation values of an already trained network when it is fed examples of a specific high level feature. Assuming the feature is isolated to a certain linear subset of the activation matrix, a vector along the decision boundary can be computed. When feeding examples, the magnitude of the directional derivative in this vectors direction can be computed on the activation matrix, yielding a measure of how much the feature is responsible for the overall derivative..ie. the total activation.
Seems like an obvious technique. A few months back, someone posted a similar linear reverse engineering technique to make a customizable face generator from already trained GANs.
- jeromebaek 8y ago"how much the feature is responsible for the overall derivative" part is obvious. But I'm curious how the assumption can be justified - how do you isolate a feature to a certain linear subset? And how do you know that this feature is (i.e.) "stripes"? The mapping from vectors to human language is the part that seems hard.
- goldenkey 8y agoBy training a binary linear classifier over the vector space of the other NNs responses to a set of inputs with that specific feature. Language shouldn't be necessary if your feature can be conveyed through examples. But yes, it's a big assumption to say that all features can be isolated as linearly decidable subsets of the activation space. I would guess one could get better results with stronger, non linear classifiers combined with more abstract generalizations to directional derivatives.