Quantization induced memory-nonlinearity transfer: Implications of analog-to-digital conversion in reservoir computing

The output-side behaviors of typical digital computing systems, such as simulated neural networks, are generally unaffected by the act of observation; however, this is not the case for the burgeoning field of physical reservoir computers (PRCs). Observer dynamics can limit or modify the natural state information of a PRC in many ways, and among the most common is the conversion from analog to digital data needed for calculations. Here, to aid in the development of novel PRCs, we investigate the effects of bounded, quantized observations on systems' natural computational abilities. By utilizing a classical reservoir computing (RC) (an echo-state network) and some PRCs (a pneumatic artificial muscle and a soft tentacle), we show that observed state quantization effectively converts a system's natural memory into higher-order, nonlinear dynamics. Furthermore, this same effect can assist in reducing detectable system errors in the presence of noise. We demonstrate how these effects, imposed only through output-end observations, can improve timer task robustness, target different computational task types, and even encode the chaotic dynamics of a Lorenz attractor in a simple linear RC in a closed loop.