Explaining noise effects in CNN: a practical case study on volcano signals
João Paulo Canário, Otávio Augusto Tessarollo Ribeiro, Ricardo A. Rios · 2022
The study presented in this work was devoted to understanding the effect of noises in classifiers created to model signals, a particular case of data analysis characterized for presenting temporal dependencies. In this sense, we have considered a real-world application related to predicting seismic volcano events based on its temporal monitoring. Firstly, we created a convolutional neural network (CNN) to model and predict such signals. Then, we have used statistical tests to confirm the noise influence on our network. Finally, we have used an eXplainable AI (XAI) approach to describe the noise behavior visually. Besides developing a CNN with an outstanding performance to deal with this real-world application, the theoretical contribution of our work is twofold: firstly, the adoption of an XAI approach, originally designed to explain image classification, to describe the influence of noise in signals; and, secondly, we show an approach to reconstruct signals after this explaining process to smooth their observations in the time domain.