Abstract

A recent, pivotal advance in the domain of neuromoprhic engineering, is the realization of reservoir computing (RC) schemes based on passive photonic components, thus merging RC’s efficiency with high bandwidth and negligible power consumption [1]. Following this direction, we demonstrate an RC scheme that relies on a phase-to-intensity mapping process, offered by the superposition of transverse modes of a large-core polymer optical fiber (POF). Randomness is provided by in-fiber defects that allow mode-coupling, whereas nonlinearity is introduced through the square-law of photodiodes (PDs) at the output. The proposed scheme is similar to [2, 3], with the exception that integrated waveguides are replaced by optical modes, while defects are operationally equivalent to couplers. Taking into consideration that defects are of wavelength-size (Mie scattering) and the number of modes can be tailored to reach 10⁶, the proposed RC can facilitate an elevated number of nodes and synapses, whereas it is also equipped with fading memory, allowing it to classify high-speed image streams.

© 2019 IEEE