This paper addresses the visualization of image classification models learned using deep Convolutional Networks (ConvNets). Two visualization techniques are introduced: one generates an image that maximizes the class score, and the other computes a class saliency map specific to a given image and class. The authors demonstrate that these methods can be used for weakly supervised object segmentation and establish a connection between gradient-based visualization and deconvolutional networks. The paper also presents a method for visualizing class models and computing image-specific class saliency maps, which can be used for object localization without additional annotation. The results show that gradient-based visualization techniques generalize the deconvolutional network reconstruction procedure.This paper addresses the visualization of image classification models learned using deep Convolutional Networks (ConvNets). Two visualization techniques are introduced: one generates an image that maximizes the class score, and the other computes a class saliency map specific to a given image and class. The authors demonstrate that these methods can be used for weakly supervised object segmentation and establish a connection between gradient-based visualization and deconvolutional networks. The paper also presents a method for visualizing class models and computing image-specific class saliency maps, which can be used for object localization without additional annotation. The results show that gradient-based visualization techniques generalize the deconvolutional network reconstruction procedure.