Alvin Kerber

math and computers and stuff

I learned some basics today about adversarial training for image recognition models. I think it’s interesting how these techniques are so generic (basically, attacking the model by introspecting its structure and/or searching its detailed output), but there hasn’t as far as I know been much discussion of these techniques in other domains. Of course, the whole concept of adversarial examples and adversarial testing is very new, but as someone who’s particularly interested in time series modeling, I’d be curious to try and see what applies in that domain. Time series is just an example; we could of course study similar situations with any complex, high-capacity-high-variance model. I bet something interesting could come out of this!