recent developments have established the utility of the use of augmentations during training of neural networks for computer vision. the intuition behind the mechanism is clear: an image close to the given one should be attributed the features close to the unperturbed one and this should lead to more robust performance whatever the task downstream.
however, and this cannot be stressed enough, the particular choice of valid augmentations depends in a very sensitive way on the particular context of the task: the notion of closeness in the image-space is not natural, but goal-dependent.
take for example a rather simple image classification problem, MNIST. while rotations by small angles, ranging from say -15° to +15° should be expected to improve performance, a rotation by 180° would transform a 6 to a 9 and vice versa, leading to confusion between the two classes.
on the other hand, a Mussolini detector could fail terribly unless rotations by up to 180° are introduced in the training set.
Δεν υπάρχουν σχόλια:
Δημοσίευση σχολίου
για συνδέστε, για συνδέστε...