to distinguish between scenes with and without tanks. Their Neural Net achieved 100% accuracy on their held out test set. When these spectacular results were presented at a conference, a person from the audience raised a concern about the training data they collected. After further investigation it turned out that all the images with tanks were taken on a cloudy day, and all images without tanks were taken on a sunny day. So, at that time the US Government was a proud owner of a multi-billion dollar computer that could tell you whether it was cloudy or not.
detectors of particular image features •We are only interested in what image features the neuron detects, not in what kind of stuff it doesnʼ’t detect •So when propagating the gradient, we set all the negative gradients to 0 •We donʼ’t care if a pixel “suppresses” a neuron somewhere along the part to our neuron CSC321: Intro to Machine Learning and Neural Networks, Winter 2016 Michael Guerzhoy
→特徴マップをGAPすれば、特徴マップにつき1つの素⼦子が 対応する(FC層の密な結合による情報の不不透明化を回避?) Learning Deep Features for Discriminative Localization Bolei Zhou, Aditya Khosla, et al., 2015