Feature selection based convolutional neural network pruning and its application in calibration modeling for NIR spectroscopy

In our previous studies, we have found that convolutional neural network (CNN) can be applied to establish calibration model in the field of near infrared (NIR) spectroscopy. However, the values of CNN parameters are carefully chosen based on trial-and-error method, including convolutional kernel width (CKW), number of convolutional kernels (NCK), stride steps etc., otherwise underfitting phenomenon may occur and the generalized performance of calibration model will become worse. The possible reason is that the relationship between these parameters and model's generalized performance is not clear. Hence, to answer this question, this paper firstly investigated the influence of these parameters in detail and found that (1) if the CNN parameters' values are not carefully designed, the number of weights between full-connected and output layer is so large that limited samples in the training set cannot well fit the nonlinear relationship. (2) while convolutional kernels move through different subintervals of NIR spectra, features extracted with varied convolutional kernel width (VCKW) are more representative for calibration modeling than with fixed convolutional kernel width (FCKW). (3) in the subintervals near those absorption peaks, little stride steps (smaller than CKW) is prefer, because it means the extracted features are overlapping, which can capture the information around absorption peaks in detail. Additionally, the experimental results also showed that generalized performance of calibration model based on extracted CNN features outperforms that of based on raw NIR spectra.