Exploring Simplified Reservoir Computing Systems for Resource-Constrained Edge AI Hardware

We present a low-overhead reservoir computing framework optimized for edge AI applications by leveraging a combination of several techniques. Specifically, we employ the Simple Cycle Reservoir (SCR) as an alternative to the widely used Echo State Network (ESN), offering a more hardware-efficient design. Our framework evaluates five simplified nonlinear activation functions alongside the conventional hyperbolic tangent function. Additionally, we employ a hyperparameter that enables a tunable trade-off between memory retention and nonlinearity by adjusting the ratio of linear to nonlinear activations. A genetic algorithm is utilized for efficient hyperparameter optimization. To further reduce hardware cost, we developed a framework for 16-bit reduced-precision arithmetic without significantly compromising model performance. Extensive evaluations on standard benchmark datasets demonstrate that our approach delivers competitive accuracy while significantly lowering computational and hardware overhead, making it well-suited for resource-constrained edge environments.