Reproduce “A Forward and Backward Compatible Framework for Few-shot Class-incremental Pill Recognition” on OpenResearch

About this paper

Jinghua Zhang, Li Liu, Kai Gao, Dewen Hu

Researchers at the National University of Defense Technology developed a framework for few-shot class-incremental pill recognition that enables continuous learning of new pill types from limited data while retaining knowledge of old ones. Their Discriminative and Bidirectional Compatible Few-Shot Class-Incremental Learning (DBC-FSCIL) framework, which includes a novel Center-Triplet loss and uncertainty-guided pseudo feature synthesis, consistently outperformed existing state-of-the-art methods, achieving an average accuracy of 92.01% on a newly introduced realistic FCPill dataset.

Key insights

• Explicitly addressing both forward compatibility (preparing for unseen classes) and backward compatibility (preventing catastrophic forgetting) within a single framework is crucial for robust few-shot class-incremental learning in fine-grained domains.
• A novel Center-Triplet (CT) loss effectively learns highly discriminative features by simultaneously enforcing intra-class compactness and inter-class separability, which is vital for fine-grained recognition.
• Generating "reliable pseudo-features" for old classes, guided by uncertainty quantification, provides an effective and storage-efficient method for data replay and knowledge distillation in incremental learning without requiring raw past data.

Summary by alphaXiv.