Visualising deep neural network decisions

Page 1. Visualization of how different patch sizes influence the result (ie, how much information is removed at once). EXPERIMENTS VISUALIZING DEEP NEURAL NETWORK DECISIONS Luisa M. Zintgraf 1, Taco S. Cohen 1, Tameem Adel1, Max Welling 1,2 1 University of Amsterdam, 2 Canadian Institute for Advanced Research We want to understand DNNs by explaining individual decisions. This can • accelerate their adoption in industry, government & healthcare, • lead to new insight and theories in poorly understood domains, • improve network architectures. [1] Robnik-Sikonja, Marko, and Igor Kononenko. “Explaining classifications for individual instances.” Knowledge and Data Engineering, IEEE Transactions on 20.5 (2008): 589-600. [2] Simonyan, Karen, Andrea Vedaldi, and Andrew Zisserman. “Deep inside convolutional networks: Visualising image classification models and saliency maps.” arXiv preprint …