Monolithically Integrated Optical Convolutional Processors on Thin Film Lithium Niobate

Photonic neural networks (PNNs) of sufficiently large physical dimensions and high operation accuracies are envisaged as ideal candidates for breaking the major bottlenecks in the current artificial intelligence architectures in terms of latency, energy efficiency, and computational power. To achieve this vision, it is of vital importance to scale up the PNNs while simultaneously reducing the high demand on the dimensions required by them. The underlying cause of this strategy is the enormous gap between the scales of photonic and electronic integrated circuits. Here, we demonstrate monolithically integrated optical convolutional processors on thin film lithium niobate (TFLN) that harness inherent parallelism in photonics to enable large-scale programmable convolution kernels and, in turn, greatly reduce the dimensions required by subsequent fully connected layers. Experimental validation achieves high classification accuracies of 96% (86%) on the MNIST (Fashion-MNIST) dataset and 84.6% on the AG News dataset while dramatically reducing the required subsequent fully connected layer dimensions to 196 × 10 (from 784 × 10) and 175 × 4 (from 800 × 4), respectively. Furthermore, our devices can be driven by commercial field-programmable gate array systems; a unique advantage in addition to their scalable channel number and kernel size. Our architecture provides a solution to build practical machine learning photonic devices.