Shouldn't it be possible to backpropogate those categorical outputs all the way back to the inputs/features (NOT weights) after a forward pass, to localize the sensitivity of them with respect to the actual pixels for a prediction? I imagine that would have to give at least some insight.
Beyond that, the convolution/max pool repeated steps could be understood to be applying something akin to a multi-level wavelet decomposition, which is pretty well understood. It's how classical matched filtering, Haar cascading, and a wide variety of proceeding image classification methods operated at their first steps too.
CNNs/Deep learning really doesn't seem like a black box at all when examined in sequence. But to me at least, randomized ensemble methods (random forest, etc.) are actually a bit more mysterious to me in their performance out of the box, with little tuning.
I'm in no way a researcher or even an enthusiast of machine learning, but I'm pretty sure that I came across a paper posted on HN a few days ago that did exactly what you and the parent poster are describing, figuring out what pixels contributed most to some machine learning algorithm. I'll try and see if I can find it.
Bagging and bootstrap ensemble methods aren't really that confusing. Just think of it as stochastic gradient descent on a much larger hypothetical data set.
The effect is same one that occurs when you get a group of people together to estimate the number of jelly beans in a jar. All the estimators are biased, but if that bias is drawn from a zero mean distribution, deviation of the average bias goes down as the number of estimators increases.
I think you might be on to something, but the big problem here is that the Input is hundreds of GB or TB's . It's hard to understand what a feature is, or even why it's selected.
I can certainly observe what's being selected once the state machine is generated, but I have no clue how it was constructed to make the features. Do determine that, I have to watch the state of the machine as it "grows" to the final result.
Beyond that, the convolution/max pool repeated steps could be understood to be applying something akin to a multi-level wavelet decomposition, which is pretty well understood. It's how classical matched filtering, Haar cascading, and a wide variety of proceeding image classification methods operated at their first steps too.
CNNs/Deep learning really doesn't seem like a black box at all when examined in sequence. But to me at least, randomized ensemble methods (random forest, etc.) are actually a bit more mysterious to me in their performance out of the box, with little tuning.