Online Program

When Do We Need a New Neural Network? (309848)

Keywords: Neural Networks, Data Depth, Online Process Monitoring, Multivariate Statistical Process Control, Low-Dimensional Representation

The rapid advancement of models based on artificial intelligence demands innovative monitoring techniques which can operate in real-time with low computational costs. If we consider machine learning algorithms that highly rely on the training dataset, it is important to guarantee that the assumption about a learned relationship between the input data and the final result is still valid during the model's deployment. If this relation holds, we can conclude that the model generates accurate predictions. Otherwise, the training or rebuilding of the model is inevitable. In neural networks, the data sample's complex features and their interactions are condensed to a multidimensional vector immediately before the output layer. In this talk, we discuss the benefit of using this compact representation of the processed data for the calculation of data depth, in order to determine the time point when retraining or revision of the model is necessary. We assess the quality of the proposed monitoring technique on neural networks with various underlying data formats, finding promising results.