alasdairtran
commented
9 years ago When it comes to choosing our training data, we might want to use the bright objects (since they have smaller errors) and then test our classifier on faint ones. But from a machine learning perspective, perhaps we might want to do the reverse, since it's better to let our classifiers know as much about the random noises at the training phase as possible.
This can be done with a robust optimisation approach.
See chapter 12 of: http://www2.isye.gatech.edu/~nemirovs/FullBookDec11.pdf