Interpreting Deep Neural Networks Towards Trustworthiness

Explore the intricacies of interpreting deep neural networks for enhanced trustworthiness in this 33-minute lecture by Bin Yu from Berkeley University, presented at the Institut des Hautes Etudes Scientifiques (IHES). Delve into the contextual decomposition (CD) method, which attributes importance to features and feature interactions for individual predictions. Discover how applying CD to interpret deep learning models in cosmology led to the development of the adaptive wavelet distillation (AWD) interpretation method. Learn about AWD's superior performance compared to deep neural networks and its interpretability in both cosmology and cell biology applications. Gain insights into the importance of quality control throughout the entire data science life cycle to build models for trustworthy interpretation.