For the past several weeks, I’ve been looking at anomaly detection using a technique called variational autoencoder (VAE) reconstruction probability. It’s a complex idea and therefore there are many possible implementations. I have designed and successfully implemented (in the sense theat I got demo programs working) several different architectures. I tried a new architecture today. I use PyTorch, which is my favorite neural code library (as opposed to Keras / Tensorflow) when I implement complex srchitectures.
In the image below, the top figure is the architecure that is implied by the original research paper that introduced the idea of VAE reconstuction probability. (The research paper is very vague regarding implementation.) There are two distributions in the architecture. My modification is shown in the bottom figure in the image. I eliminate the front-end distribution.
In the next image, I show two demo runs. The demo run on the left is from the orginal two-distribution architecture. The demo on the right uses the modified one-distribution architecture. The results are very similar, which is what I expected.
I always have several projects going on. I’ll keep plugging away at anomaly detection using VAE reconstruction probability and looking at different architecture designs. It could lead to a significant improvement in anomaly detection — or it could be a dead end. Either way, it will be interesting to explore.
Four very different cover designs for the novel “Foundation” by Issac Asimov. “Foundation” (1951) is the first of a series of three of the most famous science fiction novels ever (with “Foundation and Empire” and “Second Foundation”). “Foundation” was first written as four short stories — modular design works in fiction writing as as well as in code writing.