TiQSA: Workload Minimization in Convolutional Neural Networks Using Tile Quantization and Symmetry Approximation
Lists
Only you see this
Abstract
Convolutional Neural Networks (CNNs) in the Internet-of-Things (IoT)-based applications face stringent constraints, like limited memory capacity and energy resources due to many computations in convolution layers. In order to reduce the computational workload in these layers, this paper proposes a hybrid convolution method in conjunction with a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Particle of Swarm Convolution Layer Optimization (PSCLO)</i> algorithm. The hybrid convolution is an approximation that exploits the inherent symmetry of filter termed as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">symmetry approximation</i> and Winograd algorithm structure termed as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">tile quantization approximation</i> . PSCLO optimizes the balance between workload reduction and accuracy degradation for each convolution layer by selecting fine-tuned thresholds to control each approximation’s intensity. The proposed methods have been evaluated on ImageNet, MNIST, Fashion-MNIST, SVHN, and CIFAR-10 datasets. The proposed techniques achieved <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 5.28\text{x}$ </tex-math></inline-formula> multiplicative workload reduction without significant accuracy degradation (<0.1%) for ImageNet on ResNet-18, which is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 1.08\text{x}$ </tex-math></inline-formula> less multiplicative workload as compared to state-of-the-art Winograd CNN pruning. For LeNet, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 3.87\text{x}$ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 3.93\text{x}$ </tex-math></inline-formula> was the multiplicative workload reduction for MNIST and Fashion-MNIST datasets. The additive workload reduction was <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 2.5\text{x}$ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\sim 2.56\text{x}$ </tex-math></inline-formula> for the respective datasets. There is no significant accuracy loss for MNIST and Fashion-MNIST dataset.
BibTeX
@inproceedings{sabir2021tiqsa,
author = {Dilshad Sabir and Muhammad Abdullah Hanif and Ali Hassan and Saad Rehman and Muhammad Shafique},
title = {{TiQSA: Workload Minimization in Convolutional Neural Networks Using Tile Quantization and Symmetry Approximation}},
booktitle = {IEEE Access},
year = {2021},
url = {https://openreview.net/forum?id=ALuv_W_cxBK},
}Loading related content...