On automated source selection for transfer learning in convolutional neural networks

Muhammad Jamal Afridi, Arun Ross, Erik M. Shapiro

Research output: Research - peer-reviewArticle

Abstract

Transfer learning, or inductive transfer, refers to the transfer of knowledge from a source task to a target task. In the context of convolutional neural networks (CNNs), transfer learning can be implemented by transplanting the learned feature layers from one CNN (derived from the source task) to initialize another (for the target task). Previous research has shown that the choice of the source CNN impacts the performance of the target task. In the current literature, there is no principled way for selecting a source CNN for a given target task despite the increasing availability of pre-trained source CNNs. In this paper we investigate the possibility of automatically ranking source CNNs prior to utilizing them for a target task. In particular, we present an information theoretic framework to understand the source-target relationship and use this as a basis to derive an approach to automatically rank source CNNs in an efficient, zero-shot manner. The practical utility of the approach is thoroughly evaluated using the Places-MIT dataset, MNIST dataset and a real-world MRI database. Experimental results demonstrate the efficacy of the proposed ranking method for transfer learning.

LanguageEnglish (US)
JournalPattern Recognition
DOIs
StateAccepted/In press - 2017

Profile

Neural networks
Magnetic resonance imaging
Availability

Keywords

  • CNN selection
  • Deep learning
  • Transfer learning

ASJC Scopus subject areas

  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Artificial Intelligence

Cite this

@article{31c8db7b0f504c6aab1de3f65b58a2c6,
title = "On automated source selection for transfer learning in convolutional neural networks",
abstract = "Transfer learning, or inductive transfer, refers to the transfer of knowledge from a source task to a target task. In the context of convolutional neural networks (CNNs), transfer learning can be implemented by transplanting the learned feature layers from one CNN (derived from the source task) to initialize another (for the target task). Previous research has shown that the choice of the source CNN impacts the performance of the target task. In the current literature, there is no principled way for selecting a source CNN for a given target task despite the increasing availability of pre-trained source CNNs. In this paper we investigate the possibility of automatically ranking source CNNs prior to utilizing them for a target task. In particular, we present an information theoretic framework to understand the source-target relationship and use this as a basis to derive an approach to automatically rank source CNNs in an efficient, zero-shot manner. The practical utility of the approach is thoroughly evaluated using the Places-MIT dataset, MNIST dataset and a real-world MRI database. Experimental results demonstrate the efficacy of the proposed ranking method for transfer learning.",
keywords = "CNN selection, Deep learning, Transfer learning",
author = "Afridi, {Muhammad Jamal} and Arun Ross and Shapiro, {Erik M.}",
year = "2017",
doi = "10.1016/j.patcog.2017.07.019",
journal = "Pattern Recognition",
issn = "0031-3203",
publisher = "Elsevier Limited",

}

TY - JOUR

T1 - On automated source selection for transfer learning in convolutional neural networks

AU - Afridi,Muhammad Jamal

AU - Ross,Arun

AU - Shapiro,Erik M.

PY - 2017

Y1 - 2017

N2 - Transfer learning, or inductive transfer, refers to the transfer of knowledge from a source task to a target task. In the context of convolutional neural networks (CNNs), transfer learning can be implemented by transplanting the learned feature layers from one CNN (derived from the source task) to initialize another (for the target task). Previous research has shown that the choice of the source CNN impacts the performance of the target task. In the current literature, there is no principled way for selecting a source CNN for a given target task despite the increasing availability of pre-trained source CNNs. In this paper we investigate the possibility of automatically ranking source CNNs prior to utilizing them for a target task. In particular, we present an information theoretic framework to understand the source-target relationship and use this as a basis to derive an approach to automatically rank source CNNs in an efficient, zero-shot manner. The practical utility of the approach is thoroughly evaluated using the Places-MIT dataset, MNIST dataset and a real-world MRI database. Experimental results demonstrate the efficacy of the proposed ranking method for transfer learning.

AB - Transfer learning, or inductive transfer, refers to the transfer of knowledge from a source task to a target task. In the context of convolutional neural networks (CNNs), transfer learning can be implemented by transplanting the learned feature layers from one CNN (derived from the source task) to initialize another (for the target task). Previous research has shown that the choice of the source CNN impacts the performance of the target task. In the current literature, there is no principled way for selecting a source CNN for a given target task despite the increasing availability of pre-trained source CNNs. In this paper we investigate the possibility of automatically ranking source CNNs prior to utilizing them for a target task. In particular, we present an information theoretic framework to understand the source-target relationship and use this as a basis to derive an approach to automatically rank source CNNs in an efficient, zero-shot manner. The practical utility of the approach is thoroughly evaluated using the Places-MIT dataset, MNIST dataset and a real-world MRI database. Experimental results demonstrate the efficacy of the proposed ranking method for transfer learning.

KW - CNN selection

KW - Deep learning

KW - Transfer learning

UR - http://www.scopus.com/inward/record.url?scp=85028077324&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85028077324&partnerID=8YFLogxK

U2 - 10.1016/j.patcog.2017.07.019

DO - 10.1016/j.patcog.2017.07.019

M3 - Article

JO - Pattern Recognition

T2 - Pattern Recognition

JF - Pattern Recognition

SN - 0031-3203

ER -