Graduate School and Research Center in Digital Sciences

Semantic and visual similarities for efficient knowledge transfer in CNN training

Pascal, Lucas; Bost, Xavier; Huet, Benoit

CBMI 2019, 17th International Conference on Content-Based Multimedia Indexing, 4-6 September 2019, Dublin, Ireland

In recent years, representation learning approaches have disrupted many multimedia computing tasks. Among those approaches, deep convolutional neural networks (CNNs) have notably reached human level expertise on some constrained image classification tasks. Nonetheless, training CNNs from scratch for new task or simply new data turns out to be complex and time-consuming. Recently, transfer learning has emerged as an effective methodology for adapting pre-trained CNNs to new data and classes, by only retraining the last classification layer. This paper focuses on improving this process, in order to better transfer knowledge between CNN architectures for faster trainings in the case of fine tuning for image classification. This is achieved by combining and transfering supplementary weights, based on similarity considerations between source and target classes. The study includes a comparison between semantic and contentbased similarities, and highlights increased initial performances and training speed, along with superior long term performances when limited training samples are available.

Document Doi Hal Bibtex

Title:Semantic and visual similarities for efficient knowledge transfer in CNN training
Keywords:Transfer learning, fine tuning, image classification, model selection
Type:Conference
Language:English
City:Dublin
Country:IRELAND
Date:
Department:Data Science
Eurecom ref:5982
Copyright: © 2019 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.
Bibtex: @inproceedings{EURECOM+5982, doi = {http://dx.doi.org/10.1109/CBMI.2019.8877391}, year = {2019}, title = {{S}emantic and visual similarities for efficient knowledge transfer in {CNN} training}, author = {{P}ascal, {L}ucas and {B}ost, {X}avier and {H}uet, {B}enoit}, booktitle = {{CBMI} 2019, 17th {I}nternational {C}onference on {C}ontent-{B}ased {M}ultimedia {I}ndexing, 4-6 {S}eptember 2019, {D}ublin, {I}reland}, address = {{D}ublin, {IRELAND}}, month = {09}, url = {http://www.eurecom.fr/publication/5982} }
See also: