Towards Robust Incremental Learning under Ambiguous Supervision

TL;DR

The Prototype-Guided Disambiguation and Replay Algorithm (PGDR) is developed which leverages the class prototypes as a proxy to mitigate two intertwined challenges in IPLL, i.e., label ambiguity and catastrophic forgetting.

Abstract

Traditional Incremental Learning (IL) targets to handle sequential fully-supervised learning problems where novel classes emerge from time to time. However, due to inherent annotation uncertainty and ambiguity, collecting high-quality annotated data in a dynamic learning system can be extremely expensive. To mitigate this problem, we propose a novel weakly-supervised learning paradigm called Incremental Partial Label Learning (IPLL), where the sequentially arrived data relate to a set of candidate labels rather than the ground truth. Technically, we develop the Prototype-Guided Disambiguation and Replay Algorithm (PGDR) which leverages the class prototypes as a proxy to mitigate two intertwined challenges in IPLL, i.e., label ambiguity and catastrophic forgetting. To handle the former, PGDR encapsulates a momentum-based pseudo-labeling algorithm along with prototype-guided initialization, resulting in a balanced perception of classes. To alleviate forgetting, we develop a memory replay technique that collects well-disambiguated samples while maintaining representativeness and diversity. By jointly distilling knowledge from curated memory data, our framework exhibits a great disambiguation ability for samples of new tasks and achieves less forgetting of knowledge. Extensive experiments demonstrate that PGDR achieves superior

Figures (13)

• Figure 1: In the first task, all samples are from new classes, while subsequent tasks consist of samples containing both new and old classes. Each sample is assigned a set of candidate labels, ensuring the inclusion of the true label.
• Figure 2: (a) Class distribution based on both predicted labels and ground-truth labels. The real/estimated distribution of old classes (0$\sim$9) and new classes (10$\sim$19) in the early stage of the second task of the CIFAR100 (10 tasks). Combining the PLL baseline (PiCO) and the IL baseline (iCaRL) leads to classification bias. (b) Comparison of our PGDR and PiCO with iCaRL in IPLL. It displays the average accuracy of new and old classes for each task in CIFAR100 where PiCO with iCaRL demonstrates inferior and unstable performance compared to our method.
• Figure 3: Overall framework of PGDR. The label disambiguation module divides new and old class samples based on prototypes and assigns different labels. PGDR then combines the momentum-updated pseudo-labels to achieve label disambiguation. After completing the task, the memory replay module is utilized to filter samples for subsequent training to mitigate forgetting.
• Figure 4: Ablation experiment about disambiguation module on Tiny-ImageNet ($q$=0.2)
• Figure 5: Ablation experiment about memory replay module on Tiny-ImageNet ($q$=0.2).