An Energy Sampling Replay-Based Continual Learning Framework
Document Type
Conference Item
Publication Date
1-1-2024
Abstract
Continual Learning represents a significant challenge within the field of computer vision, primarily due to the issue of catastrophic forgetting that arises with sequential learning tasks. Among the array of strategies explored in current continual learning research, replay-based methods have shown notable effectiveness. In this paper, we introduce a novel Energy Sampling Replay-based (ESR) structure for image classification. This framework enhances the selection process of samples for replay by leveraging the energy distribution of the samples, thereby improving the effectiveness of memory samples during the replay phase and increasing accuracy. We have conducted extensive experiments across various continual learning methodologies and datasets. The results demonstrate that our approach effectively mitigates forgetting on CIFAR-10, CIFAR-100 and CIFAR-110 datasets by optimizing the replay strategy.
Keywords
Image classification, Continual learning, Catastrophic forgetting, Energy-based sampling
Divisions
fsktm,fac_eng
Funders
Universiti Malaya,Department of Artificial Intelligence, Faculty of Computer Science and Information Technology (PPRN001A-2023)
Volume
15017
Publisher
SPRINGER INTERNATIONAL PUBLISHING AG
Publisher Location
GEWERBESTRASSE 11, CHAM, CH-6330, SWITZERLAND
Event Title
Artificial Neural Networks and Machine Learning-ICANN 2024, Pt II
Event Location
Univ Italian Switzerland, Lugano, SWITZERLAND
Event Dates
17-20 September 2024
Event Type
conference
Additional Information
33rd International Conference on Artificial Neural Networks and Machine Learning (ICANN), Univ Italian Switzerland, Lugano, SWITZERLAND, SEP 17-20, 2024