Tradeoffs with physical delay feedback reservoir computing

The Reservoir Computing (RC) paradigm is a supervised machine learning scheme utilizing natural computational ability of dynamical systems. Delayed chaotic systems provide rich dynamics for information processing by using the system's transient response to an external input, so have been identified as suitable systems for reservoir computing. In delay-feedback reservoir computing a virtual network is achieved through time-multiplexing, trading off input frequency and reservoir complexity. Here we investigate several features and tradeoffs that affect physical delay-dynamical reservoirs, based on the Mackey-Glass model, including the role of input mask, and a ‘devirtualisation’ of the output. We evaluate performance using the NARMA-10 benchmark. We find that random binary masks with no offset outperform other masking schemes, but require a careful choice of model parameters. We find that reading the output of a subset of ‘virtual’ nodes, the ‘devirtualised network’, does not affect computational power, which enables a tradeoff between output design complexity and readout frequency.