Sensitivity and generalized analytical sensitivity expressions for quantitative analysis using convolutional neural networks

Abstract

In recent years, convolutional neural networks and deep neural networks have been used extensively in various fields of analytical chemistry. The use of these models for calibration tasks has been highly effective; however, few reports have been published on their properties and characteristics of analytical figures of merit. Currently, most performance measures for these types of networks only incorporate some function of prediction error. While useful, these measures are incomplete and cannot be used as an objective comparison among different models. In this report, a new method for calculating the sensitivity of any type of neural network is proposed and studied on both simulated and real datasets. Generalized analytical sensitivity is defined and calculated for neural networks as an additional figure of merit. Moreover, the dependence of convolutional neural networks on regularization dataset size is studied and compared with other conventional calibration methods.