SPIE - Education Interpreting Deep Learning Networks SC1268

Deep learning neural networks, or simply, deep neural networks (DNNs), have provided spectacular breakthroughs in the areas of artificial intelligence and machine learning with multiple applications in data science and analytics including scene recognition. One challenge faced by various end user communities is that of interpreting decisions made by DNNs. There is a prevalent notion of DNNs being "black boxes." These can be particularly confounding when erroneous decisions are made by the DNN. However, several recent investigations have proposed methods for interpreting DNNs. Methods have also been proposed for reducing the likelihood of incorrect decisions. This course will provide insights into how these methods approach the problem and into future possibilities for interpreting DNNs. Topics: • DNN examples: training & classification • The different definitions of interpretability • Interpreting weights of converged DNNs • Saliency maps in convolutional neural networks (CNNs) • Layer hierarchy and interpreting layer outputs • Interpretability vs. Explainability • Interpreting DNNs through activation analysis & deep visualization • Use of generative adversarial network models for interpretation • Correlation across layers and networks • Sensitivity analysis and Garson's algorithm • Relating classification to image features • DNN architectures to enable interpretability • Approaches for reducing likelihood of erroneous decisions through rank-N classification considerations • Future directions