Lightweight optical neural network based on micro-ring resonator

We propose a lightweight, high-efficiency, low-power optical neural network (ONN) architecture based on micro-ring resonators called the micro-ring-based depthwise separable convolution (MDSC), which achieves depthwise separable convolution of ONNs by improving the conventional broadcast-and-weight protocol structure. MDSC performs depthwise convolution using add-drop MRRs as convolution kernels to achieve efficient extraction of features for each channel of the input feature map and to minimize redundant computation across channels. The modulation phase shift of the add-drop MRRs is used as the learnable parameter, and their optical transfer functions serve as the convolution weights. To extend the number of output feature maps, the pointwise convolution stage uses trans-impedance amplifiers (TIAs) as convolution kernels and trains their magnification factors as weight values for <i>1 × 1</i> convolution. On complex network structures, MDSC achieves a decrease of 3 orders of magnitude in the number of MRRs, execution time, and energy consumption compared with traditional MRR-based ONN, with further reductions reaching up to 5 orders of magnitude as network parameters increase.