Delay-feedback reservoirs are a subset of reservoir computers characterized by a hardware-efficient architecture that trades spatial complexity for temporal processing. It employs a single non-linear node, a delay line, and a time-multiplexed input signal to generate a network of “virtual nodes,” effectively emulating a larger spatial neural network. One of the most powerful aspects of delay-feedback reservoirs is their versatility. Our previous work found that the non-linear node performs two mathematical functions, a non-linear transform and integration. The non-linear transform can be represented by any number of non-linear functions, making it difficult to optimize a delay-feedback reservoir to solve a specific computational task. This work explores different non-linear functions in order to determine their effect on the dynamics of the reservoir, in order to provide insight into this optimization problem. Five different non-linear functions are compared in terms of performance, metrics, and utilization: Mackey–Glass, sine squared, double sinusoids, Tan, and Tanh. Our results find that the Mackey–Glass non-linear function shows limited system dynamics, performing well on non-linear tasks but performing poorly on memory intensive tasks. We then demonstrate the distinct system dynamics within the other four non-linear functions. We found that sine squared shows limited overall performance, double sinusoid performs well in non-linear tasks, resembles an odd valued exponent Mackey–Glass reservoir but with greater parameter sensitivity, Tan and Tanh offers balanced performance across both task types. We find that modifying the system dynamics of a reservoir is an important step toward optimizing a delay-feedback reservoir for specific computational tasks.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)