Programmable Optical Spectrum Shapers as Computing Primitives for Accelerating Convolutional Neural Networks
Photonic convolutional accelerators have emerged as low-energy alternatives to power-demanding digital convolutional neural networks, though they often face limitations in scalability. In this work, we introduce a convolutional photonic accelerator that employs programmable kernels manifesting as trainable waveforms in the frequency domain to enable low-energy, high-throughput scalable image classification. The proposed scheme inherently provides dimensionality reduction and feature extraction directly in the optical domain. Numerical results targeting the Fashion-MNIST show that by using only 16 optical nodes, the system's classification accuracy tops at 90.1% when typical backpropagation is used. Moreover, by adapting the training technique to the forward-forward approach, a marginal drop of 1% is recorded compared to the backpropagation scenario, thus showcasing the compatibility of the overall architecture with a hardware-friendly training approach. Finally, we experimentally implement the trained kernels using a programmable waveshaper. Despite the difference between the simulated and experimentally generated transfer functions of the programmable kernels, the classification accuracy based on the experimentally obtained kernels exhibits a marginal 0.2% reduction, proving the validity of the idea and its high robustness to variations of the frequency-applied complex weights.