3D-Filtermap: A compact architecture for deep convolutional neural networks

Yingzhen Yang, Jianchao Yang, Ning Xu, Wei Han, Nebojsa Jojic, Thomas S Huang

Research output: Contribution to conferencePaper

Abstract

We present a novel and compact architecture for deep Convolutional Neural Networks (CNNs) in this paper, termed 3D-FilterMap Convolutional Neural Networks (3D-FM-CNNs). The convolution layer of 3D-FM-CNN learns a compact representation of the filters, named 3D-FilterMap, instead of a set of independent filters in the conventional convolution layer. The filters are extracted from the 3D-FilterMap as overlapping 3D submatrics with weight sharing among nearby filters, and these filters are convolved with the input to generate the output of the convolution layer for 3D-FM-CNN. Due to the weight sharing scheme, the parameter size of the 3D-FilterMap is much smaller than that of the filters to be learned in the conventional convolution layer when 3D-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 3D-FM-CNN enjoys a small parameter space by learning compact 3D-FilterMaps, while achieving performance compared to that of the baseline CNNs which learn the same number of filters as that generated by the corresponding 3D-FilterMap.

Original languageEnglish (US)
StatePublished - Jan 1 2018
Event6th International Conference on Learning Representations, ICLR 2018 - Vancouver, Canada
Duration: Apr 30 2018May 3 2018

Conference

Conference6th International Conference on Learning Representations, ICLR 2018
CountryCanada
CityVancouver
Period4/30/185/3/18

Fingerprint

neural network
Neural networks
Convolution
Neural Networks
Filter
learning
performance
Layer

ASJC Scopus subject areas

  • Education
  • Computer Science Applications
  • Linguistics and Language
  • Language and Linguistics

Cite this

Yang, Y., Yang, J., Xu, N., Han, W., Jojic, N., & Huang, T. S. (2018). 3D-Filtermap: A compact architecture for deep convolutional neural networks. Paper presented at 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada.

3D-Filtermap : A compact architecture for deep convolutional neural networks. / Yang, Yingzhen; Yang, Jianchao; Xu, Ning; Han, Wei; Jojic, Nebojsa; Huang, Thomas S.

2018. Paper presented at 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada.

Research output: Contribution to conferencePaper

Yang, Y, Yang, J, Xu, N, Han, W, Jojic, N & Huang, TS 2018, '3D-Filtermap: A compact architecture for deep convolutional neural networks', Paper presented at 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada, 4/30/18 - 5/3/18.
Yang Y, Yang J, Xu N, Han W, Jojic N, Huang TS. 3D-Filtermap: A compact architecture for deep convolutional neural networks. 2018. Paper presented at 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada.
Yang, Yingzhen ; Yang, Jianchao ; Xu, Ning ; Han, Wei ; Jojic, Nebojsa ; Huang, Thomas S. / 3D-Filtermap : A compact architecture for deep convolutional neural networks. Paper presented at 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada.
@conference{bdda0076e1704c47acd4bed8b2008320,
title = "3D-Filtermap: A compact architecture for deep convolutional neural networks",
abstract = "We present a novel and compact architecture for deep Convolutional Neural Networks (CNNs) in this paper, termed 3D-FilterMap Convolutional Neural Networks (3D-FM-CNNs). The convolution layer of 3D-FM-CNN learns a compact representation of the filters, named 3D-FilterMap, instead of a set of independent filters in the conventional convolution layer. The filters are extracted from the 3D-FilterMap as overlapping 3D submatrics with weight sharing among nearby filters, and these filters are convolved with the input to generate the output of the convolution layer for 3D-FM-CNN. Due to the weight sharing scheme, the parameter size of the 3D-FilterMap is much smaller than that of the filters to be learned in the conventional convolution layer when 3D-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 3D-FM-CNN enjoys a small parameter space by learning compact 3D-FilterMaps, while achieving performance compared to that of the baseline CNNs which learn the same number of filters as that generated by the corresponding 3D-FilterMap.",
author = "Yingzhen Yang and Jianchao Yang and Ning Xu and Wei Han and Nebojsa Jojic and Huang, {Thomas S}",
year = "2018",
month = "1",
day = "1",
language = "English (US)",
note = "6th International Conference on Learning Representations, ICLR 2018 ; Conference date: 30-04-2018 Through 03-05-2018",

}

TY - CONF

T1 - 3D-Filtermap

T2 - A compact architecture for deep convolutional neural networks

AU - Yang, Yingzhen

AU - Yang, Jianchao

AU - Xu, Ning

AU - Han, Wei

AU - Jojic, Nebojsa

AU - Huang, Thomas S

PY - 2018/1/1

Y1 - 2018/1/1

N2 - We present a novel and compact architecture for deep Convolutional Neural Networks (CNNs) in this paper, termed 3D-FilterMap Convolutional Neural Networks (3D-FM-CNNs). The convolution layer of 3D-FM-CNN learns a compact representation of the filters, named 3D-FilterMap, instead of a set of independent filters in the conventional convolution layer. The filters are extracted from the 3D-FilterMap as overlapping 3D submatrics with weight sharing among nearby filters, and these filters are convolved with the input to generate the output of the convolution layer for 3D-FM-CNN. Due to the weight sharing scheme, the parameter size of the 3D-FilterMap is much smaller than that of the filters to be learned in the conventional convolution layer when 3D-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 3D-FM-CNN enjoys a small parameter space by learning compact 3D-FilterMaps, while achieving performance compared to that of the baseline CNNs which learn the same number of filters as that generated by the corresponding 3D-FilterMap.

AB - We present a novel and compact architecture for deep Convolutional Neural Networks (CNNs) in this paper, termed 3D-FilterMap Convolutional Neural Networks (3D-FM-CNNs). The convolution layer of 3D-FM-CNN learns a compact representation of the filters, named 3D-FilterMap, instead of a set of independent filters in the conventional convolution layer. The filters are extracted from the 3D-FilterMap as overlapping 3D submatrics with weight sharing among nearby filters, and these filters are convolved with the input to generate the output of the convolution layer for 3D-FM-CNN. Due to the weight sharing scheme, the parameter size of the 3D-FilterMap is much smaller than that of the filters to be learned in the conventional convolution layer when 3D-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 3D-FM-CNN enjoys a small parameter space by learning compact 3D-FilterMaps, while achieving performance compared to that of the baseline CNNs which learn the same number of filters as that generated by the corresponding 3D-FilterMap.

UR - http://www.scopus.com/inward/record.url?scp=85070979675&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85070979675&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85070979675

ER -