Sparse-Complementary Convolution for Efficient Model Utilization on CNNs

Chun-Fu (Richard) Chen, Jinwook Oh, Quanfu Fan, Marco Pistoia, Gwo Giun (Chris) Lee

Feb 15, 2018 (modified: Oct 26, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • TL;DR: TL;DR
  • Abstract: We introduce an efficient way to increase the accuracy of convolution neural networks (CNNs) based on high model utilization without increasing any computational complexity. The proposed sparse-complementary convolution replaces regular convolution with sparse and complementary shapes of kernels, covering the same receptive field. By the nature of deep learning, high model utilization of a CNN can be achieved with more simpler kernels rather than fewer complex kernels. This simple but insightful model reuses of recent network architectures, ResNet and DenseNet, can provide better accuracy for most classification tasks (CIFAR-10/100 and ImageNet) compared to their baseline models. By simply replacing the convolution of a CNN with our sparse-complementary convolution, at the same FLOPs and parameters, we can improve top-1 accuracy on ImageNet by 0.33% and 0.18% for ResNet-101 and ResNet-152, respectively. A similar accuracy improvement could be gained by increasing the number of layers in those networks by ~1.5x.
  • Keywords: CNN, sparse convolution, sparse kernel, sparsity, model utilization, image classification
0 Replies