Some New Machine Learning Methods for Materials Data Analysis

Quan Qian*, Yan Feng

Shanghai University, Shanghai, 200444, China

EXTENDED ABSTRACT: Data driven materials R&D has been an important part of the materials genome engineering. In this talk, we will introduce some new machine learning methods, such as multi-task learning (MTL), multi-modal learning(MML), and privacy preserving machine learning(PPML), etc. Specifically, for MTL, firstly, the attention mechanism is used to extract the similarity among different samples, then using the cross stitch network to share the information between different tasks, and finally the weighted comprehensive loss function is used to coordinate the training progress of each task. For MML, we integrate multi-modal data, for instance, text, image, video, to train a comprehensive model at the same time, which can digging the data from different view or point. For PPML, aiming at protection the privacy and security of materials data (sometimes just sensitive attributes of the data), a secure multi-party aggregation machine learning model is studied. The model can not only protect local sensitive data, but also protect the network gradient information during machine learning of each client. A hierarchical grid security aggregation model is proposed, where homomorphic encryption and secret sharing can protect local data from external attacks. Meanwhile, the hierarchical grid structure can provide different trust levels to prevent internal attacks. Also the proxy server in different layers can share the load pressure on the server side, so as to reduce service congestion, memory overflow and such abnormal states. From several practical materials data experiments, it shows that our methods not only has good prediction effect, but also can preserve the sensitive data privacy.