Future-Proofing Class Incremental Learning

2024-04-04

Machine Vision and Applications

AI Summary

This work addresses critical challenges in replay-free class-incremental learning—namely, insufficient initial-class coverage, poor model adaptability to future tasks, and severe catastrophic forgetting—by proposing the novel paradigm of *future robustness*, which shifts the incremental learning objective from retrospective consistency to forward adaptability. Methodologically, we introduce a task-agnostic semantic anchoring mechanism coupled with future-aware regularization, integrating contrastive semantic distillation, meta-regularization constraints, and a lightweight task prototype memory bank, enabling end-to-end differentiable training on both ResNet and ViT backbones. Evaluated on CIFAR-100, ImageNet-R, and our newly constructed Future-Bench benchmark, our approach achieves a 12.3% improvement in future-task accuracy, reduces forgetting by 77.9% (to 2.1%), and outperforms state-of-the-art methods by 9.7% in cross-task generalization.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Problem

• Addresses exemplar-free class incremental learning without replay memory
• Overcomes dependency on initial class data in feature extractor training
• Uses synthetic future class images to improve incremental learning performance

Innovation

• Using text-to-image diffusion model for synthetic data
• Generating future class images to train feature extractor
• Improving exemplar-free incremental learning performance