High-dimension Photonic Multi-Channel Convolution Parallel Acceleration Unit

In increasingly complex machine learning tasks, multiple convolutional kernels are essential for extracting diverse features from data. The parallel computation of these kernels becomes crucial as tasks grow in complexity. Optical neural networks offer significant acceleration for these computations, leveraging the high bandwidth and low latency advantages. In this work, we design a novel photonic multi-channel convolution acceleration unit, capable of performing parallel computations of multiple convolutional kernels within a single clock cycle. This design achieves an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$M$</tex>-fold increase in parallelism, where <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$M$</tex> represents the number of convolutional kernels being used, compared to single-kernel computation. We constructed an optical experimental system to validate the reliability of this architecture. The system utilizes radio frequency, wavelength, and spatial multiplexing to encode information onto the amplitude of radio frequency signals and map convolutional kernels to the power of optical signals and represent the size of the convolutional kernels by the number of spatial modulators, enabling photonic multi-channel parallel convolution functionality. This setup achieved a 97% accuracy rate in handwritten digit recognition tasks. Additionally, our high-dimension processing architecture is flexible and scalable, allowing for adjustments the number and values of convolution kernels according to the specific requirements of different tasks.