A CNN Approximation Method Based on Low-bit Quantization and Random Forests
Authors:
- 8 June 2021
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 039-044
Abstract
In recent years, the use of image recognition technology in edge devices has been increasing. To achieve low-power and low-latency inference of convolutional neural networks (CNNs) in edge devices, methods that reduce the number of operations, such as pruning, have been actively researched. However, even after applying these existing methods, we still need to calculate many multiply-accumulate (MAC) operations. In this paper, we propose a hardware-friendly CNN approximation method based on low-bit quantization and random forests to reduce the number of operations and operation cost of CNN inference. In our experiments, we reduce the number of operations by 30.8% for LeNet and by 27.1% for ResNet18 while maintaining high image classification accuracy.Keywords
Funding Information
- Technology Development
This publication has 5 references indexed in Scilit:
- FBNA: A Fully Binarized Neural Network Accelerator2018
- A Threshold Neuron Pruning for a Binarized Deep Neural Network on an FPGAIEICE Transactions on Information and Systems, 2018
- Deep Residual Learning for Image Recognition2016
- Backpropagation Applied to Handwritten Zip Code RecognitionNeural Computation, 1989
- Induction of decision treesMachine Learning, 1986