Integrated Photonic-electronic Deep Neural Networks: from Sub-nanosecond Image Classification to PVT-tolerant Activation Functions

Photonic deep neural networks can perform fast and energy-efficient classification through computation by propagation and low-loss interconnects. Here, a review of our work on the first integrated end-to-end photonic-electronic deep neural network for direct image classification that achieved a classification time of 570 ps is presented. Then, the first monolithically integrated, reconfigurable, and trainable optical nonlinear activation function (NLAF) designed in a 45nm CMOS SOI process for scalable silicon photonic deep neural networks is introduced. The NLAF supports reconfigurable ReLU and Sigmoid functionalities over a 6 GHz of bandwidth, with trainable parameters for turn-on thresholds, slopes, and saturation levels to effectively mitigate process, voltage, and temperature (PVT) variations, ensuring robustness in deep networks by reducing the number of trained parameters by orders of magnitude. The chip was implemented using a 45nm CMOS SOI process within a footprint of 0.028 mm² and achieves a minimum activation power of -23 dBm.