Why do CNNs generalize so poorly to small image transformations?


New top story on Hacker News: Why do CNNs generalize so poorly to small image transformations?