Integrated Optical Neural Network Processor

Artificial intelligence research has experienced several booms and busts, and the current AI technology surge since 2012 is mostly the result of three factors: (1) the invention of deep learning algorithms. (2) the growth of massive data on the Internet. (3) the development of AI hardware, in particular, the introduction of Nvidia CUDA. 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 and 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 thus discover a novel way to enhance computing power. According to studies, optical has high bandwidth, low latency (light speed), low energy consumption, and multiple forms of multiplexing (bosonic properties). One study indicated that optical neural networks are 103 faster in speed and 10−2 lower in power consumption compared to digital circuits. Consequently, optical solutions are a feasible alternative. Some representative optical neural networks were created by the researchers, including spiking neural networks, reservoir computing, diffractive neural networks, coherent MZI networks, and broadcast and weight neural networks. However, the first three optical neural networks focus on specific algorithms or scenarios. The latter two approaches are based on photonic integrated circuit platforms and focus on general matrix multiplication, which is the core operation in neural networks for deep learning. The WDM-MRM broadcast neural network is superior to the coherent MZI network due to the fact that MRM has a smaller feature size than MZI, operates with parallel connections, has superior optical loss performance, and avoids SVD decomposition. Analyzing the architecture of latter two solutions reveals that they are analog solutions, which are inherently sensitive to noise, and requiring DACs to convert digital electrical signals to analog signals for input into the optical system. In addition, the output signal is analog, requiring a high-speed and high-precision ADC to process it. Therefore, in order to increase the accuracy of an analog computing system, the system’s signal-to-noise ratio must be increased. Alternatively, we can switch the optical neural network from an analog to a digital system, which has a better noise tolerance. In particular, in systems with low signal-to-noise ratio, such as the high-speed case, the digital signal system can still obtain a high accuracy result compared to the analog system. Therefore, we propose a novel architecture: the digital optical neural network (DONN), which is based on the broadcast and weight system. Unlike broadcast and weight architecture, the DONN scheme uses the heater modulator in MRM as the input for the weight value and the high-speed RF port as the input for the feature matrix value. Since the DONN scheme focuses on the inference process with fixed weights, it is possible to keep the weight matrix values in the analog state while leaving the feature matrix input values in the digital state, resulting in a PAM signal output that is easier to detect and more precise than the analog signal. In the DONN scheme, the number of PAM signal levels increases with the number of input channels; however, research and experience indicate that PAM levels greater than 16 are difficult to implement at frequencies beyond 40 GHz. Therefore, a higher-order format than PAM is required to represent more states. 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. After digitizing the input feature matrix values and increasing the dimensionality of the output signal format, we discovered that imprecise control of the weight values and the inherent noise of photodetector would lead to results with low precision or error. Therefore, this research work 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 number of ”convolution kernels” without increasing the scale of the optical hardware. With the DONN scheme and the aforementioned improved methods, this study implements a high-speed, low-power, low-latency and high-precision optical neural network based on a photonic integrated circuit platform for general-purpose matrix multiplication processing.