Learning $3$D-FilterMap for Deep Convolutional Neural Networks
Abstract
We present a novel and compact architecture for deep Convolutional Neural Networks (CNNs) in this paper, termed D-FilterMap Convolutional Neural Networks (D-FM-CNNs). The convolution layer of D-FM-CNN learns a compact representation of the filters, named D-FilterMap, instead of a set of independent filters in the conventional convolution layer. The filters are extracted from the D-FilterMap as overlapping D submatrics with weight sharing among nearby filters, and these filters are convolved with the input to generate the output of the convolution layer for D-FM-CNN. Due to the weight sharing scheme, the parameter size of the D-FilterMap is much smaller than that of the filters to be learned in the conventional convolution layer when D-FilterMap generates the same number of filters. Our work is fundamentally different from the network compression literature that reduces the size of a learned large network in the sense that a small network is directly learned from scratch. Experimental results demonstrate that D-FM-CNN enjoys a small parameter space by learning compact D-FilterMaps, while achieving performance compared to that of the baseline CNNs which learn the same number of filters as that generated by the corresponding D-FilterMap.
Keywords
Cite
@article{arxiv.1801.01609,
title = {Learning $3$D-FilterMap for Deep Convolutional Neural Networks},
author = {Yingzhen Yang and Jianchao Yang and Ning Xu and Wei Han},
journal= {arXiv preprint arXiv:1801.01609},
year = {2018}
}