A Convolutional Neural Network Solution for Spectroscopic Data Regression

This work deals with the problem of predicting chemical information from different materials using spectroscopic data. This is part of a field of study called chemometrics, which combines chemistry with informatics. In pattern recognition, this kind of problem is known as multivariate regression. In this work, we propose a convolutional neural network (CNN) that combines global and local features of the spectroscopic signal. The motivation behind this method is that convolutional layers in CNN provide localized features only because the filters have a limited width (such as 3×3 or 5×5). However, global features are also important in learning the regression function. The proposed CNN is composed of two branches one branch learns global features from the signal while the second branch learns local features using convolutional layers. The two branches are combined at the end of the deep network using a concatenation operation. The preliminary results presented on two chemometric datasets show clearly the potential of the proposed deep learning method.