In 2012 Krizhevsky et al. demonstrated that a Convolutional Neural Network (CNN) trained on large amount of data as part of the ImageNet challenge significantly outperformed traditional computer vision approaches on image classification. Subsequently, Girshick et al. (2013) exploited these features to establish the new state of the art on PASCAL object detection. This suggests that computer vision maybe in the process of a feature revolution akin to that following SIFT and HOG nearly a decade ago. It is therefore important to get more insights into features learned by these networks. Our work provides answer to the following four questions: (a) What happens during finetuning of a discriminatively pretrained network? (b) How much information is in the location and how much of it is in the magnitude of filter activation? (c) Does a multilayer CNN contain Grand-Mother Cells? How distributed is the code? (d) How does training of CNN progress over time? Does too much pre-training hurt generalization performance?




Download Full History