Analytical Calculation of Partial Derivatives in Deep Neural Networks

This project will research on efficient methods for calculating the partial derivatives of output variables with respect to input variables in at least one of the following deep neural network architectures: CNN, RNN, GNN. Partial derivatives are crucial for understanding the sensitivity of neural networks to changes in input variables and are widely used in various applications, including optimization, sensitivity analysis, and model interpretability. While numerical differentiation methods have been traditionally employed, this research project aims to explore analytical approaches to calculate partial derivatives in deep neural networks. The project will investigate the effectiveness, accuracy, and computational efficiency of different analytical techniques, expanding the current methods existing in the NeuralSens R library or the neuralsens Python package.