ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Retraining Image Classification Model with Ensemble Networks and Episodic Memories in Three Continual Learning Scenarios
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
Soonyong Song, Heechul Bae, Hyonyoung Han, Youngsung Son
Issue Date
2020-06
Citation
Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 2020, pp.1-4
Language
English
Type
Conference Paper
Abstract
In this competition, we evaluated three continual learning (CL) scenarios with a CORe50 dataset. To preserve past knowledge from catastrophic forgetting, we trained classification model through multi-head and latent replay approaches. To increase classification performance, an ensemble model was implemented in feature network composed of three excellent backbone networks. Also, early stopping and learning rate scheduling schemes were applied to maximize prediction accuracy without overfitting. Our approaches showed score 0.89 by average with respect to the three scenarios in this qualification step.
KSP Keywords
Backbone Network, Catastrophic forgetting, Classification Performance, Classification models, Early stopping, Ensemble models, Ensemble networks, Image Classification, Learning rate, Learning scenarios, Multi-head