Wavelength-multiplexed Delayed Inputs for Memory Enhancement of Microring-based Reservoir Computing

Among their benefits is the potential for parallel processing using mature technologies such as wavelength division multiplexing (WDM) which has been applied to photonic neural networks [2]. In our previous study [3], we showed the potential of WDM to improve the computing capabilities of a reservoir computing (RC) scheme. RC is a relatively recent computing paradigm that uses random fixed weights and complex nonlinear dynamics to map the input data to a higher dimensional space. This process allows simplifying its training process, which is only required in its output layer (usually through linear or ridge regression). The performance of RC is highly related to the nonlinear dynamics of the nodes and the capability to buffer past inputs (memory) [1]. Nevertheless, photonic RC schemes may reduce their scalability if multiple photonic blocks are implemented as nonlinear nodes.