Modular DFR: Digital Delayed Feedback Reservoir Model for Enhancing Design Flexibility

In RC, the reservoir weights are not altered and the weights of the output layer that follows the reservoir are the target of learning [13], which allows efficient training. RC is considered suitable for time series processing because of its recurrent structure, which reflects past inputs. Because the weights of a reservoir are fixed, it can be implemented in hardware utilizing various physical phenomena. A delayed feedback reservoir (DFR) [2] is a specific type of RC system. It is particularly suitable for hardware implementations because it can be compactly constructed with a single nonlinear element and a feedback loop [18]. Until now, hardware implementations of DFRs have been of two types: analog and digital [2]. In an analog implementation, only the nonlinear element of the reservoir or the nonlinear element and the feedback loop are implemented in an analog manner. However, the inputs and outputs are generally processed digitally, requiring a digital-toanalog converter (DAC) and an analog-to-digital converter (ADC). In addition, the time required for signal propagation through the feedback loop reduces the throughput.