Let's say we have simple pictures dataset containing 40x40 images of digits. We have only one image of each digit. We want to use that as training set, but we need more data, so we use data augmentation. We use only simple operations like translating and rotating and generate 1000 more images of each digit.


Natural way to do data augmentation will be randomly generating parameters like translate_x, translate_y and rotate and applying them into our base image.

Does distribution of these parameters matter? From one hand we would like to have net that is able to recognize digit placed in the side of the image and rotated as well as digit placed in center and not rotated at all but maybe we don't need such accuracy with those borderline images? Maybe we know that our prediction data will be close to the centered ones so we want high accuracy in that cases and the more digit is translated and rotated the lower our's net accuracy might be.

What I mean is can we augmented data with parameters with e.g. gaussian distribution to make our net more sensitive for cases closest to ideal and less sensitive to this borderline cases. Advantage of that would be less training data that we don't need and more control on characteristic of our neural net.


This digits case is just simple example to show You what I mean.



You must log in to answer this question.