A high-performance optical computing accelerator for regression algorithm training

The regression algorithm is one of the most important tasks for numerical prediction and classification in the field of artificial intelligence. Its training process depends on a large number of matrix operations, which bring great challenges to performance and power consumption. Optical computing, a new computing paradigm, is expected to solve the dilemma of regression algorithm training with its advantages of ultra-low power consumption and high parallelism. Herein, a bidirectional optical computing accelerator based on the Mach-Zehnder interferometer network is proposed, which uses the optical signal on a single integrated photonic chip to propagate in the forward and backward directions respectively to realize the matrix operation in parallel at high speed to complete the training of the regression algorithm. Specifically, it consists of three steps: 1) Separate the matrix-vector multiplication module from the gradient information; 2) Design a bidirectional optical computing architecture to implement the above matrix operation module; 3) Complete the electronic-photonic converging computing. Numerical experiments of three typical tasks, logistic regression, linear regression, and ridge regression, verify the feasibility of the bidirectional optical computing accelerator to achieve the above training of the regression algorithm. Theoretical research indicates that when performing regression algorithm training at 100 GHz data detection rate, our scheme achieves 2 orders of magnitude computational speed and 2 orders of magnitude power efficiency compared to conventional electronic computer architectures.