Fig. 3: Implementation with racetrack resonators.
From: Fully nonlinear neuromorphic computing with linear wave scattering
a, Neuron modes are represented by racetrack resonators (light blue); racetrack resonators of different layers in the neural network are crossed, employing techniques to reduce the cross-talk between them. They can either be coupled via tunable couplings, or smaller racetrack resonators with tunable detunings (dark blue)—the coupler modes—which change the effective coupling as the detuning is varied. This is illustrated in the ersatz image (right). b, The advantage of this design is that the system can be scaled up and requires only a minimal number of waveguide crossings; the neuron modes can still be accessed with waveguides from the outside. c, Two possibilities to measure the gradients: following the expression for the gradient in terms of the scattering matrix elements, one can either measure the response at the racetrack resonators and use equation (8) or, if coupling modes are used, directly at the coupling resonators and use equation (7) to update the parameters. As an alternative to coupling waveguides to each resonator, optical grating tap monitors can be utilized that light up according to the output signal at the resonator, which can be recorded by a camera15,50. Grating tap monitors can be complemented by integrated photodetectors for a faster readout. To be sensitive to a specific quadrature (equation (4)), the light coupled to the grating tap monitor can be combined with light from a local oscillator (not shown) to perform a homodyne measurement. d, Scale of the relevant frequencies in an optical implementation. e, Distribution of neuron-mode detunings and tunable couplings after training.
