Benchmarking Reservoir and Recurrent Neural Networks for Human State and Activity Recognition

Abstract

Monitoring of human states from streams of sensor data is an appealing applicative area for Recurrent Neural Network (RNN) models. In such a scenario, Echo State Network (ESN) models from the Reservoir Computing paradigm can represent good candidates due to the efficient training algorithms, which, compared to fully trainable RNNs, definitely ease embedding on edge devices. In this paper, we provide an experimental analysis aimed at assessing the performance of ESNs on tasks of human state and activity recognition, in both shallow and deep setups. Our analysis is conducted in comparison with vanilla RNNs, Long Short-Term Memory, Gated Recurrent Units, and their deep variations. Our empirical results on several datasets clearly indicate that, despite their simplicity, ESNs are able to achieve a level of accuracy that is competitive with those models that require full adaptation of the parameters. From a broader perspective, our analysis also points out that recurrent networks can be a first choice for the class of tasks under consideration, in particular in their deep and gated variants.