A Divide-and-Conquer Approach towards Understanding Deep Networks

• DOI: 10.48550/arxiv.1907.06194
• Publication Date: 2019-01-01
• Authors: Weilin Fu; Katharina Breininger; Roman Schaffert; Nishant Ravikumar; Andreas Maier
• Keywords: Interpretability; Black box; Deep Neural Networks

Deep neural networks have achieved tremendous success in various fields including medical image segmentation. However, they long been criticized for being a black-box, that interpretation, understanding and correcting architectures is difficult as there no general theory deep network design. Previously, precision learning was proposed to fuse traditional approaches. constructed this way benefit from the original known operator, fewer parameters, improved interpretability. do not yield state-of-the-art performance all applications. In paper, we propose analyze using operators, by adopting divide-and-conquer strategy replace components, whilst retaining its performance. The task of retinal vessel segmentation investigated purpose. We start with high-performance U-Net show step-by-step conversion are able divide into modules operators. results indicate combination trainable guided filter version Frangi yields at level (AUC 0.974 vs. 0.972) reduction parameters (111,536 9,575). addition, trained layers can be mapped back their algorithmic interpretation analyzed standard tools signal processing.