PhD defence by Xiansong Meng

Title: Integrated Optical Neural Network Processor

Supervisors
Principal supervisor: Senior Researcher Hao Hu, Department of Electrical and Photonics Engineering, DTU, Denmark
Co-supervisor: Senior Researcher Yunhong Ding, Department of Electrical and Photonics Engineering, DTU, Denmark
Co-supervisor: Researcher Deming Kong, Department of Electrical and Photonics Engineering, DTU, Denmark

Evaluation Board
Associate Professor Michael Galili, Department of Electrical and Photonics Engineering, DTU, Denmark
Professor Christophe Peucheret, University of Rennes 1 - FOTON Institute, France
Department Director Arindam Mallik, Interuniversity Microelectronics Centre (IMEC), Belgium

Master of the Ceremony
Senior Researcher Francesco Da Ros, Department of Electrical and Photonics Engineering, DTU, Denmark

Abstract
The calculation hardware evolution contributes to the current AI technology surge since 2012. According to the research, the demand for computing power is doubling every three to four months due to the development of AI technology. As a result, computing hardware has evolved from CPUs, GPUs to FPGAs and AISCs. However, these advances in electronic hardware are based on the evolution of Moore's Law, which is coming to an end. To satisfy the increased demand for the development of AI technologies, we must discover a novel way to enhance computing power. The photon with large bandwidth, low latency, low energy consumption and multiple forms of multiplexing has demonstrated the huge advantage in neural network processing. However, after summarizing the state of the art of optical neural networks, we found that they are all analog schemes, which are inherently sensitive to noise and cannot process high precision tasks. Therefore, to increase the accuracy of an analog computing system, the system's signal-to-noise ratio must be increased. Alternatively, we propose a DONN (Digital Optical Neural Network) scheme to achieve high computational precision under applied noise scenarios. 

By digitizing the input matrix data based on the MRM (Micro Ring Modulators) broad and cast scheme, DONN achieves large-scale integration and high-precision computation to process general matrix multiplication, which is the core operation in the AI computation process. The calculation precision is comparable with the computer result. However, as the number of channels increases, the output PAM level increases dramatically. Therefore, the QAM constellation based on IQ modulation is introduced to replace the PAM output format to accommodate more output states and further reduce the SNR requirement of the optical computing system. In addition, the high frequency computing system based on QAM-DONN scheme has no signal distortion caused by DC effect, compared to the baseband computing system of DONN scheme. Although, the input data has been digitized and the output is also a PAM or QAM signal that can be easily detected. However, the weighting matrix remains in the analog state. The imprecise control of the weight values and the inherent noise of photodetector can lead to low accuracy or errors in the results. Therefore, this research attempts to digitize the weight values utilizing the LUT (look-up table) technique. By incorporating the LUT method into the electrical part, we implement matrix multiplication with arbitrary weight values that can simultaneously process a large amount of "convolution kernels" without increasing the scale of the optical hardware.

With the DONN scheme and the improved method described above, this study implements a high-speed, low-power consumption, low-latency and high-precision optical neural network based on a photonic integrated circuit platform for general-purpose matrix multiplication processing.