Toward Knowledge Extraction in Classification of Volcano-Seismic Events: Visualizing Hidden States in Recurrent Neural Networks

Understanding how deep hierarchical models build their knowledge is a key issue in the usage of artificial intelligence to interpret the reality behind data. Depending on the discipline and models used, such knowledge may be represented in ways that are more or less intelligible for humans, limiting further improvements on the performance of the existing models. In order to delve into the characterization and modeling of volcano-seismic signals, this article emphasizes the idea of deciphering what and how recurrent neural networks (RNNs) model, and how this knowledge can be used to improve data interpretation. The key to accomplishing these objectives is both analyzing the hidden state dynamics associated with their hidden units as well as pruning/trimming based on the specialization of neurons. In this article, we process, analyze, and visualize the hidden states activation maps of two RNN architectures when managing different types of volcano-seismic events. As a result, the class-dependent discriminative behavior of most active neurons is analyzed, thereby increasing the comprehension of the detection and classification tasks. A representative dataset from the deception island volcano (Antarctica), containing volcano-tectonic earthquakes, long period events, volcanic tremors, and hybrid events, is used to train the models. Experimental analysis shows how neural activity and its associated specialization skills change depending on the architecture chosen and the type of event analyzed.