Silvia Scarpetta — Dept. of Physics "E.R.Caianiello" University of Salerno, Italy #
Ensemble of Convolutional Neural Networks mitigating overconfident predictions of seismic traces. #
(Giovanni Messuti, Silvia Scarpetta, Ortensia Amoroso,Ferdinando Napolitano, Mariarosaria Falanga,Paolo Capuano) Deep Neural Networks (DNNs) demonstrated remarkable power in various domains, achieving state of the art performance in several tasks such as image recognition, natural language processing and so on. DNNs often face significant challenges with out-of-sample data, where the model encounters data points that differ substantially from the training set, producing unjustified overconfident predictions that lead to poor generalization. Ensemble learning is a technique where multiple models are trained to solve the same problem. Aggregating the predictions of several models, the ensemble can achieve better generalization and robustness than any individual model. Ensembles are also known to enhance uncertainty estimation, offering more reliable confidence intervals and mitigating the risks associated with out-of-sample data. In our study, we build an ensemble of Convolutional Neural Networks to classify seismic traces based on first motion polarity, aggregating the predictions of individual networks through Unweighted Model Averaging. Our ensemble model demonstrates a greater ability to manage out-of-sample data, mitigating the effect of overconfident predictions. As a result, noise-only waveforms and waveforms lacking polarity information are correctly distinguished from waveforms containing polarity information.