Binary decision diagram-based binary neural network quantitative analysis method

Abstract

The invention relates to a binary decision diagram-based binary neural network quantitative analysis method, and provides a binary decision diagram-based BNN network coding mode by means of a binary decision diagram (BDD) and by analyzing the internal structure characteristics of a binary neural network. According to the method, the internal structure of the neural network is fully utilized, the input and output relation of the neural network is converted into a cardinality constraint set with the module as the unit, BDD coding is conducted on the obtained constraint set through the binary decision diagram, and then BDD coding of the whole neural network is completed. Compared with an existing quantitative analysis scheme capable of proving approximation, the method is more efficient and accurate, incremental coding of the neural network is supported, and the robustness and interpretability of the neural network can be accurately analyzed. According to the method, not only is the precision greatly improved, but also the performance is better, and the scale of the network which can be processed by the method is far larger than that of other BDD-based coding methods.

Description

technical field [0001] The invention relates to a neural network technology, in particular to a binary neural network quantitative analysis method based on a binary decision graph. Background technique [0002] At present, deep neural network technology is increasingly incorporated into various application fields, such as autonomous driving and medical diagnosis. Modern neural networks usually contain a large number of parameters, which are usually stored as 32 / 64 bit floating point numbers, and require a lot of floating point operations to calculate the output of a single input. Therefore, deploying them on resource-constrained embedded devices is often a challenge. To alleviate this problem, quantization techniques emerge as a promising technique. In particular, 1-bit quantized binary neural networks (BNNs) can not only reduce memory storage overhead, but also compute outputs by performing bit operations, thereby greatly reducing runtime and improving energy efficiency. ...

AI Technical Summary

Problems solved by technology

[0005] At present, the research results of quantitative analysis of quantized deep neural networks are very limited. For binary neural networks, existing methods either only support small-scale BNNs, or can only provide provably approximate results. To achieve higher accuracy and Confidence, verification costs are often prohibitively high

Images