All-Photonic Artificial Neural Network Processor Via Non-linear Optics

Integrated photonic architectures have been shown to accelerate conventional computing tasks that have been deemed as bottlenecks in traditional electronics. Among these tasks, one that is fundamental to neural networks is matrix processing. We propose the use of an all-optical integrated chip to accelerate deep learning. The architecture we introduce encodes information in the complex amplitudes of frequency modes that act as neurons. Information processing by intermodulating the neurons is implemented by the nonlinear optical process of Four-Wave Mixing (FWM). The FWM among neurons and controlled pump modes allows us to implement the linear transformation within microring resonators. Our nonlinear activation function relies on pulse distortion via nonlinear interactions, followed by controlled capture of these pulses into microring resonators. The proposed design is novel in its ability to attain arbitrarily large computational speeds by increasing the power of the pump modes, within limits imposed by heating due to losses. Additionally, it provides a completely unitary weight matrix, thus opening up the prospects of reversible computing. Through simulations, we show that our design achieves performances commensurate with present-day techniques on classification benchmarks.