Inverting Convolutional Networks with Convolutional Networks

Deep representations, in particular ones implemented by convolutional neural networks, have led to good progress on many learning problems. However, the learned representations are hard to analyze and interpret, even when they are extracted from visual data. We propose a new approach to study deep image representations by inverting them with an up-convolutional neural network. Application of this method to a deep network trained on ImageNet provides numerous insights into the properties of the feature representation. Most strikingly, the colors and the rough contours of an input image can be reconstructed from activations in higher network layers and even from the predicted class probabilities.