Paper Info
Title
Convolution Neural Network Development Support System using Approximation Methods to Evaluate Inference Accuracy and Memory Usage in an Embedded System.
Abstract
Convolution neural networks have become widely used in embedded systems such as automatic driving systems. In these cases, inference functions in convolution neural networks are implemented due to resource limitation in embedded systems. Field-programmable gate array implementation is preferable because of low power consumption and real-time response time in embedded systems. Parameters in a convolution neural network are floating-point numbers, and enormous floating-point calculation is required. This is a challenge because the field-programmable gate array has a limited floating processing unit and in-memory processing. One way to solve the problem is to approximate an enormous number of parameters and to perform efficient computation. In the approximation, floating-point numbers of parameters are implemented using smaller-size integers or numbers having fewer bits. However, the inference accuracy decreases in the approximation, leading to a tradeoff situation. In addition, which layer should be approximated in order to be effective is not clear. In order to solve these problems, we developed an approximation support system. The developed system approximates the parameters and calculates the accuracy of the parameters and the required memory size. Furthermore, using this system, we carry out experiments to evaluate the effectiveness of several approximation methods for a large-scale network and dataset.
Year
DOI
Venue
2019
10.1109/SmartWorld-UIC-ATC-SCALCOM-IOP-SCI.2019.00242
SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI
DocType
Citations 
PageRank 
Conference
0
0.34
References 
Authors
0
5
Name
Order
Citations
PageRank
Toki Matsumoto100.34
Yukikazu Nakamoto27921.50
Ryota Yamamoto311.39
Shinya Honda400.34
Kazutoshi Wakabayashi500.34