CN-115759222-B - Photon neural network system based on multitasking neural network

CN115759222BCN 115759222 BCN115759222 BCN 115759222BCN-115759222-B

Abstract

The invention provides a photonic neural network system based on a multi-task neural network, which comprises an optical calculation module, a multi-task neural network module and an FPGA-based optical calculation communication and control module, wherein the optical calculation module takes a photonic artificial intelligent chip as a core and modulates optical signals to complete multiplication, summation and nonlinear operation of optical analog quantities with high precision and high speed, the multi-task neural network module is provided with a main branch neural network and a plurality of secondary branch neural networks and can process a plurality of classification tasks and regression tasks at the same time, and the optical calculation and control module based on the FPGA is connected with the optical calculation module and the multi-task neural network module and is used for realizing real-time high-speed communication of the optical calculation module and the multi-task neural network module. The system has the characteristics of high bandwidth and low energy consumption, greatly reduces the time of serial operation of the original neural network by utilizing high-dimensional parallel calculation, has the advantages of large network scale and high parallel operation speed, and can improve the recognition accuracy and shorten the training time on the basis of single task.

Inventors

ZHAO JIAN
SU ZIHAO
XU JINSHENG

Assignees

天津大学

Dates

Publication Date: 20260508
Application Date: 20221129

Claims (8)

1. A photonic neural network system based on a multi-tasking neural network, comprising: The optical calculation module takes the photon artificial intelligent chip as a core and is used for loading information of optical signals, modulating the optical signals and completing multiplication, summation and nonlinear operation of high-precision and high-speed optical analog quantities; the multi-task neural network module is provided with a main neural network and a plurality of secondary neural networks which are added by taking the main neural network as a main line according to tasks, can process a plurality of classification tasks and regression tasks at the same time, extracts weight information of the network, and transmits the weight information into the optical calculation communication and control module based on the FPGA; the optical calculation communication and control module based on the FPGA is connected with the optical calculation module and the multi-task neural network module and is used for realizing real-time high-speed communication between the optical calculation module and the multi-task neural network module and realizing real-time updating of the weight of the optical calculation module; The optical calculation module comprises a laser, a photon artificial intelligent chip and a photoelectric converter, wherein the laser transmitter is connected with the optical calculation communication and control module based on the FPGA, the laser transmitter is controlled to emit light pulses according to information received from the optical calculation communication and control module based on the FPGA, the laser transmitter is connected with the photon artificial intelligent chip by adopting an optical fiber, continuous light input is provided for the photon artificial intelligent chip, and the photon artificial intelligent chip is used for modulating light signals input by the optical transmitter; And the secondary neural network of the multi-task neural network module provides corresponding loss functions and optimization functions for different tasks, and adjusts the preset multi-task neural network model to finish corresponding task functions.
2. The photonic neural network system of claim 1, wherein the photonic artificial intelligence chip is comprised of a cascade of mach-zehnder interferometer arrays.
3. The photonic neural network system of claim 2, wherein the photonic artificial intelligence chip is configured to perform multiplication, summation, and nonlinear operations over the optical domain.
4. The photonic neural network system of claim 2, wherein the photonic artificial intelligence chip employs a multidimensional multiplexing technique to achieve parallel transmission in a spatial dimension in the design of the photonic neural network.
5. The photonic neural network system of claim 1, wherein the main neural network of the multi-tasking neural network module is configured to extract features for each task identification.
6. The photonic neural network system of claim 1, wherein in the multi-tasking neural network module, the training weights of the primary neural network may be randomly initialized or preset during training, and the training weights of the primary neural network are smaller than the training weights of the secondary neural network.
7. The photonic neural network system of claim 1, wherein the FPGA-based optical computing communication and control module comprises an FPGA chip, an AD/DA module, an input/output interface module, a clock module, and a memory module, wherein: The FPGA chip is used for processing the input signals according to the preset parallelism to complete real-time high-speed communication between the optical calculation module and the neural network; The AD/DA module is connected with the FPGA chip, the clock module and the input/output interface module, provides four-channel AD/DA conversion and is used for digital-to-analog conversion between the optical calculation module and the FPGA; The input/output interface module is connected with the FPGA chip and the AD/DA module and is used for carrying out interface expansion on the FPGA chip and providing a high-speed PCIe 3.0 interface, a network port and an SFP optical fiber interface; The clock module is connected with the FPGA chip module and the AD/DA module and is used for providing a working clock for the FPGA chip; and the memory module is connected with the FPGA chip and used for exchanging storage data with the FPGA.
8. The photonic neural network system of any one of claims 1-7, wherein the photonic neural network system is a optoelectric hybrid integrated board card, the optical computation module forms an optical part, and the multi-tasking neural network module forms an electrical part with the FPGA-based optical computation communication and control module.

Description

Photon neural network system based on multitasking neural network Technical Field The invention relates to the technical field of artificial intelligence, in particular to a photonic neural network system based on a multi-task neural network. Background At present, the calculation power required by the neural network algorithm model is mainly provided by a GPU server and an electronic neural network chip, the electronic chip is limited by moore's law, and the updating and iteration cycle of the electronic chip needs 12-18 months and cannot keep pace with the increase of the calculation power demand. Furthermore, the electronic chip has von neumann bottleneck, namely, the hardware framework of the electronic chip causes the neural network to read and move the stored data back and forth when running, thereby adding additional energy consumption and time cost. The photonic neural network based on the simulation framework operation can successfully avoid von Neumann bottleneck in the optical field, has the advantage of high bandwidth which is not possessed by electronic signals, and can fully utilize the parallel processing capability of light to solve the problem of the electronic neural network. With the development of technology, a deep learning mode is more and more popular, and compared with a traditional manual feature classification mode, the deep learning method improves classification accuracy and saves manpower and material resources. Compared with the traditional single-task method, the multi-task deep learning method has higher recognition precision, shortens training time and improves efficiency. Disclosure of Invention The object of the present invention is to address the deficiencies of the prior art and to provide a photonic neural network system based on a multi-tasking neural network. The system reduces the influence of the limitation of the electronic chip on the operation speed of the traditional neural network, and utilizes the photon artificial intelligent chip to realize the high-speed parallel, low-power consumption and miniaturized neural network. The invention provides a photonic neural network system based on a multi-task neural network, which comprises an optical calculation module, a data processing module and a data processing module, wherein the optical calculation module takes a photonic artificial intelligent chip as a core and is used for loading information of optical signals, modulating the optical signals and completing multiplication, summation and nonlinear operation of high-precision and high-speed optical analog quantities; the multi-task neural network module is provided with a main neural network and a plurality of secondary neural networks which are added by taking the main neural network as a main line according to tasks, and is used for processing the digital electric signals preprocessed by the FPGA-based optical calculation communication and control module, simultaneously processing a plurality of classification tasks and regression tasks, extracting the weight information of the network and transmitting the weight information into the FPGA-based optical calculation communication and control module; and the optical calculation communication and control module based on the FPGA is connected with the optical calculation module and the multi-task neural network module and is used for realizing real-time high-speed communication between the optical calculation module and the multi-task neural network module and realizing real-time updating of the weight of the optical calculation module. The invention uses the photon artificial intelligent chip to form the optical calculation module, which expands the bandwidth, reduces the energy consumption, avoids the heating problem and breaks through the limitation of moore's law compared with the existing electronic chip. According to the invention, by using a multidimensional multiplexing technology, in the design of the photonic neural network, different light pulses transmitted in a single time dimension are realized by using a space dimension, so that the time of serial operation of the original neural network is greatly reduced, and the advantages of large network scale and high parallel operation speed are realized. The invention fully utilizes the advantage of multi-task deep learning, performs characteristic sharing among a plurality of tasks, improves the recognition precision on the basis of a single task, and simultaneously trains the plurality of tasks, thereby effectively shortening the training time and improving the efficiency. Drawings FIG. 1 is a system block diagram of the present invention; FIG. 2 is a system block diagram of a light calculation module; FIG. 3 is a block diagram of a multi-tasking neural network; FIG. 4 is a system diagram of an FPGA-based optical computing communication and control module; Detailed Description The technical solutions of the embodiments of the present invention will be clearly