We welcome our new visiting overseas student Zijing Luo from China University of Geosciences. She is in Room 410, Madsen Building and you are welcome to drop by and say hello. Her research project is “Interpretable deep learning algorithm for Mineral Prospectively Mapping”. Below is her research project summary.
Deep learning (DL) algorithms have been proven to be powerful tools for mining complex and non-linear geospatial data and extracting unknown patterns of geological processes. In mineral exploration, DL methods have been adopted to enhance the extraction of mineralization information and recognize hidden patterns (Zuo et al., 2019). Despite the excellent performance of DL in mineral prospectivity mapping (MPM), their nested nonlinear network structures and deep abstraction expression make such models highly opaque (Castelvecchi, 2016; Samek et al., 2017; Reichstein et al., 2019). This opaque black-box property prevents validating, interpreting, and understanding the reasoning of the model (Reichstein et al., 2019). As a result, the decision results of end-to-end DL models cannot be fully trusted. Understanding the meaning behind the features automatically extracted by a DL algorithm, and explaining the decisions made by a DL algorithm are from current research hotspots (Ribeiro et al., 2016; Karpatne et al., 2017a; Raissi, 2018). In addition, although the data-driven DL algorithm can increase the accuracy of model output through multiple trainings, it may not accurately model the metallogenic regularity due to the lack of geological information constraints. This means that even if the model performs well on the train set or the test set, it may have large deviations in situations outside the valid domain (Karpatne et al., 2017a, 2017b). That is, there is a phenomenon that is inconsistent with the actual geological prior knowledge (Reichstein et al., 2019). My research project focuses on two approaches involving both post-hoc interpretability analysis and ad-hoc interpretable modeling to improve the interpretability of the DL algorithm applied to MPM. And the metallogenic model and DL algorithm are combined to make the delineated metallogenic prospect area consistent with geological knowledge.
