Restricted Boltzmann Machine method for dimensionality reduction of large spectroscopic data

Multivariate data obtained using, for instance, Laser-Induced Breakdown Spectroscopy (LIBS), are quite bulky and complex. Advanced processing of spectroscopic data demands a multidisciplinary approach, covering not only modern machine learning tools but also a deep understanding of underlying physical mechanisms. Dimension reduction and visualization of large datasets is a task of significant interest in the spectroscopic data processing. Commonly employed linear techniques (e.g., Principal Component Analysis, PCA) cannot explain the correlations of higher-order which are present in the data. Even more, computational cost and memory limitations become way more relevant considering the size of "modern" LIBS data (millions of high-dimensional spectra). Methods based on Artificial Neural Networks (ANN) seem suitable for this task, and based on their success, they are given considerable attention within the spectroscopic community. We propose a new methodology based on Restricted Boltzmann Machine (ANN method) for dimensionality reduction of spectroscopic data and compare it to standard PCA. As an extension to successful reconstruction, we demonstrate a generation of new (unseen) spectra by the RBM model trained on a large spectroscopic dataset. This data generation is of great use not only for the extending measured datasets but also as a proper training state's confirmation of the model.