Deep Inside Convolutional Networks: Visualising Image Classification\n Models and Saliency Maps

Abstract

This paper addresses the visualisation of image classification models, learnt\nusing deep Convolutional Networks (ConvNets). We consider two visualisation\ntechniques, based on computing the gradient of the class score with respect to\nthe input image. The first one generates an image, which maximises the class\nscore [Erhan et al., 2009], thus visualising the notion of the class, captured\nby a ConvNet. The second technique computes a class saliency map, specific to a\ngiven image and class. We show that such maps can be employed for weakly\nsupervised object segmentation using classification ConvNets. Finally, we\nestablish the connection between the gradient-based ConvNet visualisation\nmethods and deconvolutional networks [Zeiler et al., 2013].\n