Photonic Extreme Learning with Chaotic Microcavities

Recent years have seen a surge in the demand for application-specific neuromorphic hardware, alongside a growing applications space. In the context of optical neural networks, integrated Photonic Extreme Learning Machines (PELMs) offer key advantages for hardware implementation, while meeting requirements for scalability, energy-efficiency and operation bandwidth. PELMs are photonic neural networks with feed-forward connectivity whose operation exploits the intrinsically complex input-to-output information transformation of the network, with training only applied via the readout layer [1]. Chaotic microcavities, i.e. dielectric microcavities with geometries that sustain aperiodic ray trajectories, have been proposed for use as compact physical reservoirs [2], [3]. Previous studies have focussed on numerical modelling of the microcavities, where the wave propagation dynamics can be time-resolved and play the role of a short-term memory. In experimental implementations, the ultrashort time scales involved in the optical path propagation are unpractical to detect and a more efficient mode of operation of such microcavities is as PELMs in classification tasks. Here, we demonstrate fabrication of silicon nitride (SbN4) based chaotic resonators and show application of these as the hidden layer of a PELM for image classification.