• bayesian optimization
• machine learning
• echo state networks
• recurrent neural networks
• biomedical applications

In recent years, we are witnessing a shift in paradigm in how and where Machine Learning (ML) inference is performed, moving to embedded devices with limited computational capabilities.

The Reservoir Computing paradigm proposes a very efficient approach for tackling time-centric problems. In the case of Echo State Networks (ESN), a sparse, untrained non-linear network is paired with a trained readout layer, leaving a much greater importance to the configuration of the hyper-parameters of the system.

We propose EchoBay, a C++ library for ESN design and training. EchoBay aims to achieve maximum performance on different devices, searching for the optimal ESN configuration for each case study. This can be done thanks to the Bayesian Optimization (BO) process, which
automatically searches hyper-parameters that maximize a fitness function.

This system is validated on benchmark and real-world tasks, such as the prediction of blood pressure in a non-invasive way.
Considering embedded devices, we show that BO is able to take in account the limited memory or computational resources of these devices.
In the blood pressure tasks, our solution is comparable with clinical recommended devices (according to the British Hypertension Society).
Finally, we propose different metrics to evaluate the quality of the reservoir in unsupervised and semi-supervised settings, showing that they are able to guide the BO towards optimal configurations, compared to the supervised case.