Rapid detection and identification of chemical hazards at low concentrations can save lives. Spectroscopy techniques, such as Surface-Enhanced Raman Scattering (SERS), have attracted lots of attention for the detection of molecules. SERS relies on the inelastic scattering of excitation light and molecular resonance to generate molecular fingerprint maps. Each map contains unique information about a molecule. SERS is powerful as it is a real-time detection method and can quickly and effectively detect a variety of chemical structures and material composition. To identify components in the analyte with the spectroscopy data, a conventional two-stage model usually consists of data preprocessing and modeling. Misuse of preprocessing may introduce artifacts or remove informative patterns and result in worse model performance. Therefore, developing a model that can be optimized end-to-end towards the final goal of identifying components is important.
Advanced machine learning algorithms and deep neural networks have demonstrated great success in numerous applications. However, since the performance of these algorithms highly depends on the amount of the available training data, a standard scenario is to centralize the optimization of the models as well as the collected data. This centralized training may compromise user privacy and data confidentiality, which can be especially troublesome when the data are from the health or commercial sectors. An emerging machine learning approach to handle these issues is called federated machine learning.
This project aims to enable fast, accurate, and robust detection of chemical hazards with federated machine learning. It will investigate to what degree a fully automated algorithm can rapidly identify constituent components in the analyte, quantify the uncertainty in its prediction, and be optimized and deployed in federated cloud computing.