On the Post-hoc Explainability of Deep Echo State Networks for Time Series Forecasting, Image and Video Classification

Since their inception, learning techniques under the Reservoir Computing\nparadigm have shown a great modeling capability for recurrent systems without\nthe computing overheads required for other approaches. Among them, different\nflavors of echo state networks have attracted many stares through time, mainly\ndue to the simplicity and computational efficiency of their learning algorithm.\nHowever, these advantages do not compensate for the fact that echo state\nnetworks remain as black-box models whose decisions cannot be easily explained\nto the general audience. This work addresses this issue by conducting an\nexplainability study of Echo State Networks when applied to learning tasks with\ntime series, image and video data. Specifically, the study proposes three\ndifferent techniques capable of eliciting understandable information about the\nknowledge grasped by these recurrent models, namely, potential memory, temporal\npatterns and pixel absence effect. Potential memory addresses questions related\nto the effect of the reservoir size in the capability of the model to store\ntemporal information, whereas temporal patterns unveils the recurrent\nrelationships captured by the model over time. Finally, pixel absence effect\nattempts at evaluating the effect of the absence of a given pixel when the echo\nstate network model is used for image and video classification. We showcase the\nbenefits of our proposed suite of techniques over three different domains of\napplicability: time series modeling, image and, for the first time in the\nrelated literature, video classification. Our results reveal that the proposed\ntechniques not only allow for a informed understanding of the way these models\nwork, but also serve as diagnostic tools capable of detecting issues inherited\nfrom data (e.g. presence of hidden bias).\n