The silicon-based integrated DONN chips introduced in Chap. 3 are all designed for single tasks. Although they have advantages such as low latency and low power consumption in specific application scenarios, they still have limitations. First, the input signal loading is implemented using thermo-optic phase shifters, which means they cannot handle tasks with high-speed signal inputs. Second, they cannot deal with multiple tasks, limiting the applications of silicon-based integrated DONN chips. Third, they cannot directly handle tasks with serial input signals. Therefore, the main research content of this chapter is to design a multifunctional silicon-based integrated DONN chip. The specific arrangements are as follows: Sect. 4.1 will introduce the architecture of the multifunctional silicon-based integrated DONN network; Sect. 4.2 will introduce the serial-to-parallel conversion technology based on arrayed waveguide structures, which is a key method for realizing high-speed serial signal input; Sect. 4.3 will introduce the silicon-based integrated DONN designed based on multi-task learning and transfer learning methods, and verify the performance of a single silicon-based integrated DONN in solving multiple ten-classification tasks through numerical calculations. Finally, Sect. 4.4 will summarize the content of this chapter.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Research on Multifunctional Silicon-Based Integrated Diffractive Optical Neural Network

  • Tingzhao Fu

摘要

The silicon-based integrated DONN chips introduced in Chap. 3 are all designed for single tasks. Although they have advantages such as low latency and low power consumption in specific application scenarios, they still have limitations. First, the input signal loading is implemented using thermo-optic phase shifters, which means they cannot handle tasks with high-speed signal inputs. Second, they cannot deal with multiple tasks, limiting the applications of silicon-based integrated DONN chips. Third, they cannot directly handle tasks with serial input signals. Therefore, the main research content of this chapter is to design a multifunctional silicon-based integrated DONN chip. The specific arrangements are as follows: Sect. 4.1 will introduce the architecture of the multifunctional silicon-based integrated DONN network; Sect. 4.2 will introduce the serial-to-parallel conversion technology based on arrayed waveguide structures, which is a key method for realizing high-speed serial signal input; Sect. 4.3 will introduce the silicon-based integrated DONN designed based on multi-task learning and transfer learning methods, and verify the performance of a single silicon-based integrated DONN in solving multiple ten-classification tasks through numerical calculations. Finally, Sect. 4.4 will summarize the content of this chapter.