Prediction of residuals

[gngdb]

I can't find a good link for this so I'm going to try to explain it here. If we take the predictions of the Pylearn2 model on a validation set we can take the difference between the probability that it predicts and the label (encoded 1-of-k). If our model were perfect this would always be zero.

We'll see more variation in these residuals when our model predicts badly. Therefore, we should be able to use these residuals as the target of a multivariate regression problem and use the features from OpenCV to predict them. Don't know what regressor to use; maybe a Random Forest or similar?

If the OpenCV features are good at this task then it's likely that they are going to be useful to improve our score. If not, then they're not necessarily going to be useful to us, but they might still be.