Rollback Ensemble With Multiple Local Minima in Fine-Tuning Deep Learning Networks
- Authors
- Ro, Y.; Choi, J.; Heo, B.; Choi, J.Y.
- Issue Date
- Sep-2022
- Publisher
- Institute of Electrical and Electronics Engineers Inc.
- Keywords
- Deep neural network; fine-tuning; Generative adversarial networks; Image retrieval; image retrieval; learning strategy; Learning systems; Neural networks; person reidentification.; Task analysis; Training; Training data
- Citation
- IEEE Transactions on Neural Networks and Learning Systems, v.33, no.9, pp 4648 - 4660
- Pages
- 13
- Journal Title
- IEEE Transactions on Neural Networks and Learning Systems
- Volume
- 33
- Number
- 9
- Start Page
- 4648
- End Page
- 4660
- URI
- https://scholarworks.bwise.kr/cau/handle/2019.sw.cau/62555
- DOI
- 10.1109/TNNLS.2021.3059669
- ISSN
- 2162-237X
2162-2388
- Abstract
- Image retrieval is a challenging problem that requires learning generalized features enough to identify untrained classes, even with very few classwise training samples. In this article, to obtain generalized features further in learning retrieval data sets, we propose a novel fine-tuning method of pretrained deep networks. In the retrieval task, we discovered a phenomenon in which the loss reduction in fine-tuning deep networks is stagnated, even while weights are largely updated. To escape from the stagnated state, we propose a new fine-tuning strategy to roll back some of the weights to the pretrained values. The rollback scheme is observed to drive the learning path to a gentle basin that provides more generalized features than a sharp basin. In addition, we propose a multihead ensemble structure to create synergy among multiple local minima obtained by our rollback scheme. Experimental results show that the proposed learning method significantly improves generalization performance, achieving state-of-the-art performance on the Inshop and SOP data sets. IEEE
- Files in This Item
- There are no files associated with this item.
- Appears in
Collections - Graduate School of Advanced Imaging Sciences, Multimedia and Film > Department of Imaging Science and Arts > 1. Journal Articles
![qrcode](https://api.qrserver.com/v1/create-qr-code/?size=55x55&data=https://scholarworks.bwise.kr/cau/handle/2019.sw.cau/62555)
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.