Reducing Reservoir Dimensionality with Phase Space Construction for Simplified Hardware Implementation

Abstract

In this paper, we propose a One-Shot AI algorithm employing reservoir computing in combination with Phase Space Reconstruction (PSR). Such a combination not only simplifies the reservoir structure-by reducing both the number of neurons and the complexity of neural interconnections-thereby facilitating hardware implementation but also maintains high accuracy. When applied to gait authentication for 72 subjects, our approach achieves an impressively low Equal Error Rate (EER) of 0.08% with just 64 neurons required in the reservoir. Notably, these neurons do not need to be interconnected, enhancing the algorithm's suitability for hardware implementation. Furthermore, the energy efficiency of our PSR-based reservoir computing algorithm is significantly enhanced by two main strategies: 1) Only one device at each time step is activated, and 2) Adjustable intervals for reading out device observables.