CEAT: Continual Expansion and Absorption Transformer for Non-Exemplar Class-Incremental Learning
Core Concepts
CEAT proposes a novel architecture for Non-Exemplar Class-Incremental Learning, addressing plasticity-stability dilemma and classifier bias.
Abstract
Abstract:
Real-world applications require models to learn new tasks continuously without forgetting old knowledge.
CEAT architecture extends expanded-fusion layers in parallel with frozen parameters to learn novel knowledge.
Prototype contrastive loss reduces overlap between old and new classes in feature space.
Introduction:
Class Incremental Learning aims to recognize new classes without catastrophic forgetting.
NECIL prohibits storing old images or using pre-trained models.
Methodology:
CEAT consists of continual expansion and absorption steps to maintain parameter number after each task.
Batch interpolation pseudo-feature generation maintains decision boundaries of previous classes.
Prototype contrastive loss enforces inter-class separation.
Experiment:
Tested on CIFAR100, TinyImageNet, ImageNet-Subset benchmarks.
CEAT outperforms previous works with significant improvements in accuracy.
Related Work:
Previous methods store exemplars from previous tasks, while NECIL focuses on non-exemplar learning.
CEAT
Stats
CEAT achieves 5.38%, 5.20%, and 4.92% improvement on CIFAR-100, TinyImageNet, and ImageNet-Subset benchmarks respectively.
Quotes
"Experience-Replay methods store a subset of the old images for joint training."
"To address the classifier bias towards the new classes, we propose a novel approach to generate the pseudo-features."