Deconvolution Network - rugbyprof/5443-Data-Mining GitHub Wiki

Deconvolutional networks (De Conv Nets) have been proposed to visualize image patterns that strongly activate any given neuron in a Convolutional Neural Networks and therefore shed some light on the Convolutional Neural Networks structure. However, the De Conv Net construction is partially heuristic and so are the corresponding visualizations. Main aim of these networks is to learn hierarchy of features in an unsupervised manner.Now its being used to invert the downsampling that takes place in a convolutional network and expand the image back to its original size.

Reference: https://www.robots.ox.ac.uk/~vedaldi/assets/pubs/mahendran16salient.pdf