Simple and fast back propagation and training algorithm of binary neural network

Abstract

The invention designs a simple and fast back propagation and reverse mirror image training algorithm of the binary neural network, and the input and output matrix weight values of the binary neural network are solved by combining reverse mirror image training. According to the algorithm, only simple bit operation, bit flipping operation and para-bit matching operation are used, compared with a binary neural network algorithm, the calculation amount is saved, and meanwhile hardware implementation is facilitated. The neural network model can be explained, and behaviors of the neural network model are easy to pre-judge. And the neural network model design is facilitated. And meanwhile, the algorithm is small in training sample size and good in robustness. The training time is saved. The algorithm is applied to various artificial intelligence deep learning fields requiring a neural network, especially to the edge computing field with limited computing resources, and to the deep learning field requiring real-time and dynamic learning.

Description

Technical field: [0001] artificial intelligence deep learning Background technique: [0002] At present, the back-propagation algorithm of binary neural network is complex. [0003] At present, the binary neural network does not have a reverse mirror training method, and the training samples are large and the convergence is slow. Invention content: [0004] The invention designs a simple and fast back-propagation and reverse mirror training algorithm of the binary neural network, which is used for the fast training of the binary neural network and the application of the binary neural network model. [0005] It is characterized by: [0006] 1. There are only simple bit operations, bit flip operations, and bit matching operations. [0007] 2. A simple back-propagation algorithm that uses a simple bit flip operation to obtain the input and output matrix weights. [0008] 3. Reverse mirror training method, fast convergence. [0009] 4. Compared with the general binary neu...

AI Technical Summary

Problems solved by technology

[0002] At present, the backpropagation algorithm of binary neural network is complex

[0003] At present, there is no reverse mirror training method for binary neural networks, the training samples are large, and the convergence is slow

Images