I recently came across batch-instance normalization in one of my classes. The idea is rather clever: it is supposed to help with removing unnecessary styles on, say, an image, in order to help with neural style transfer and similar tasks. From what I understand, It is basically a linear combination of instance normalization and batch normalization. [Paper here] (https://arxiv.org/abs/1805.07925)
Intuitively, it seems as if this idea would help in more generic classification tasks as well by improving generalization performance over a dataset that might not be completely uniform in its ‘style’. It is also easily implemented and thus could make a good addition to the fast.ai library after being properly tested.
Thoughts?