Hardware-Application Co-Design to Evaluate the Performance of an STDP-based Reservoir Computer

Reservoir computer (RC) is an emerging computing framework to optimize the training cost. RC is a suitable solution for low-power devices such as edge devices. In addition, RC layer is a mystery box to the researcher. Usually, a very small neuron size and minimal random connectivity are used for the RC layer. However, different percentages of connectivity and neuron size can influence the performance of RC. In addition, when the neuron size is greater and the connectivity is higher the RC with spike-timing-dependent plasticity (STDP) can show better performance with energy overhead. In this work, the RC layer is modified based on different connectivity, neuron size, and adaptation of STDP. This different configuration is evaluated with various applications. Usually, a small-size RC layer and less connectivity show better performance without STDP. On the other hand, higher connectivity and a larger RC layer show better performance with STDP adaptation. About 24% accuracy is increased with connectivity scaling and approximately 8% performance is enhanced with STDP adaptation higher activity overhead.