FPGA system and implementation method based on on-line training neural network of quasi-newton method

Abstract

The invention discloses a FPGA system and implementation method based on an on-line training neural network of the quasi-newton method. The FPGA system comprises modules of LS, GC, HU, NNE, CSC and PNG. The implementation method includes 1 analyzing c++ codes of the quasi-newton method, dividing the algorithm into three calculation modules, and converting each calculation module into a hardware block by editing Verilog; 2 determining the hardware structure of a neural network evaluation module NNE by editing Verilog according to the topological structure, training method and excitation function of the neural network; 3 generating a module PNG by realizing the random number on the basis of a 32-bit linear shift register; 4 adopting a FPGA on-chip memory as a buffer to link the five hardware modules, storing the middle calculation results, and determining the operation order of the five modules and implementing data delivery between memory and corresponded modules in a manner of a finite state machine; 5 carrying out performance test through hardware design. The speed of neural network training can be increased through the FPGA, and the real-time capability requirements of neural network on-line training can be met.

Description

technical field [0001] The invention relates to the field of accelerated design of FPGA, in particular to an FPGA system and an implementation method for online training neural network based on quasi-Newton method. Background technique [0002] A neural network is an information processing system that has the ability to learn arbitrary input-output relationships from a set of data. Training is a key step in establishing the neural network structure. The traditional offline training method has the following problems: 1) If the sample data is time-varying, it cannot accurately capture the relationship between input and output. 2) If all samples are trained, the training speed is slow and may fall into local optimum (such as literature [1]). Therefore, in recent years, the combination of neural network and online training methods has been widely used in signal processing, speech recognition, sequence prediction (such as literature [2]) and other fields. [0003] Currently, o...

AI Technical Summary

Problems solved by technology

The traditional offline training method has the following problems: 1) If the sample data is time-varying, it cannot accurately capture the relationship between input and output

2) If all samples are trained, the training speed is slow and may fall into local optimum (such as literature [1])

[0004] GPU is widely recognized as one of the optional accelerators, but its high power consumption is the Achilles' heel of embedded applications (such as literature [4])

Images