Mitigate forgetting in few-shot class-incremental learning using different image views

Open Webpage

Pratik Mazumder, Pravendra Singh

Published: 2023, Last Modified: 25 Jan 2026Neural Networks 2023EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Demonstrates that impact of forgetting is different for different views of same image.•Proposes Augmentation-based Prediction Rectification (APR) approach for FSCIL.•Proposes Feature Synthesis Module (FSM) for FSCIL.•Significantly reduces catastrophic forgetting in the FSCIL setting.•Augments and improves the performance of other FSCIL methods.