Implementation of Anomaly Detection with VGG-16 model

[Danelrf]

Hello and thank you for the great research!

I was wondering if it is possible to find an implementation of the anomaly detection KDE explanation used in the paper

"Explaining the Predictions of Unsupervised Learning Models" from the book "xxAI - Beyond Explainable AI"

There it is exemplified as identifying an anomalous slit within a picture of wood, trained with the VGG16 network.

In any case, thank you as well for the K-Means demo.

[jacobkauffmann]

Hi @Danelrf

The easiest way to get these results is to use d(a, uk) = a @ (a - uk).detach() - (uk @ (a - uk)).detach() as a distance function, and then use the gradient of $o(a)$ w.r.t. a. This gives you the relevance in VGG16 feature space.
You can then use zennit to propagate relevance through VGG16.

We are also working on implementing more elements in zennit directly (such as distance, LogMeanExp, and so on). I will come back at you, then.