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

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models face catastrophic forgetting in few-shot class incremental learning (FSCIL), losing prior knowledge when learning new classes with limited data.
• Data augmentation is commonly used to improve model performance in FSCIL, but its effectiveness can be limited by how augmented views activate model neurons.

Purpose of the Study:

• To investigate how different data augmentation views affect neuron activation and information storage in FSCIL models.
• To develop novel methods to mitigate catastrophic forgetting and improve performance in the FSCIL setting.

Main Methods:

• Proposed an Augmentation-based Prediction Rectification (APR) approach to reduce catastrophic forgetting by utilizing diverse augmented data views.
• Introduced a novel Feature Synthesis Module (FSM) to generate relevant features for previously seen classes without requiring original training data.

Main Results:

• Demonstrated that information from different augmented views is stored in distinct model neurons, offering a pathway to mitigate forgetting.
• APR significantly improved the performance of existing FSCIL approaches.
• FSM outperformed other generative methods for feature synthesis in the FSCIL context.

Conclusions:

• The proposed APR and FSM methods effectively reduce catastrophic forgetting in FSCIL.
• The findings highlight the importance of considering data augmentation strategies for preserving knowledge in incremental learning.
• The developed approach shows superior performance on benchmark datasets compared to existing methods.