This article is based on Alex's CNN code, which uses visualization techniques to bring the features learned from each layer of convolutional neural networks to a human-visible, feature visualization, and tries to propose improvements. is equivalent to the inverse process of convolutional neural networks.
The main frameworks are as follows:The main use of the technology has unpooling, rectification,filtering (inverse filter) The main analytical processes are:1.Architecture Selectionproblem found: The first layer lters is a mix of extremely and low frequency information, with Litt Le coverage of the mid frequencies. Additionally, the 2nd layer visualization shows aliasing artifacts caused by the large Stride 4 used in The 1st layer convolutions. (Though I don't know how he sees it (⊙o⊙)) 2.Occlusion SensitivityGet rid of a part and see what happens.3.Correspndence AnalysisCorrelation Analysis
From for notes (Wiz)
Visualing and understanding convolutional networks