Input reduction of convolutional neural networks with global sensitivity analysis as a data-centric approach

Pruning methods are used for dealing with the rapid growth of neural network parameters as the neural network develops. These enable a reduction in not only the size of the network, but also the bandwidth it utilizes. In this article, global sensitivity analysis methods, like Sobol and eFAST, are applied to determine the least significant inputs. Data-centric Artificial Intelligence is a fledgeling science and is gathering more and more attention among top researchers. This methodology is focused on data manipulation and constant model architecture, while the model-centric approach modifies models. This article cuts off itself from typical model-centric pruning methods and is focused solely on data pruning. In the work, principal component analysis is applied as a benchmark for the proposed Global Sensitivity methods. This work is focused on convolutional neural networks. It shows how the accuracy of CNNs is impacted by input reduction – doing so through the example of three image datasets and three series datasets.