High-quality activation function for applications in neuromorphic photonic chips realized using low-quality nonlinear optical resonators

TL;DR

This paper introduces a novel nonlinear switching mechanism in low-quality optical resonators that enables efficient, tunable, and sharp threshold functions for neuromorphic photonic chips, even with weak nonlinearities and conventional materials.

Contribution

It proposes a new switching mechanism in nonlinear resonators that achieves sharp, tunable thresholds independent of resonator quality, suitable for neuromorphic photonic applications.

Findings

Achieves sharp threshold behavior near critical coupling regime.

Allows threshold broadening control via variable losses.

Operates effectively with weak Kerr nonlinearities and conventional materials.

Abstract

Integrated optical devices that can realize a threshold filtration of signals are in demand in photonics. In particular, they play a key role in neuromorphic chips, acting as optical neurons. A list of requirements exist for thresholders to be practically applicable in this context. A value of the threshold, in general, should be independently tunable for each neuron. A sharpness of the corresponding step function should be also dynamically variable, to allow switching between deterministic and stochastic algorithms, for example, in optical Ising machines. Nonlinear ring resonators draw the attention of researchers in this field since they potentially can provide the required type of threshold behavior. Here we suggest the switching mechanism that implements the property of resonators to provide extremely sharp (with respect to the wavelength) $\pi$ phase shifts near the critical coupling regime. Adding a variable source of losses into the resonator, a well controlled broadening of the threshold can be achieved. Sharpness of the step in this case is independent on the quality factor of resonators and corresponding width of resonances. The nonlinear switching mechanism allows to use this feature to construct efficient optical neurons, that operate at small intensities. It also allows to use conventional materials like silicon with a relatively weak Kerr nonlinearity. The obtained results potentially lead the way to fast nonlinear Kerr effect based activation functions that can operate in continuous wave regime.