Class Incremental Learning via Contrastive Complementary Augmentation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Class incremental learning (CIL) aims for continuous knowledge acquisition from data streams while preserving prior knowledge.
• Existing CIL methods face performance degradation due to imbalanced old and new data, hindering the balance between learning new information and retaining old knowledge.

Purpose of the Study:

• To propose a novel method, Contrastive Complementary Augmentation Learning (CoLA), to address the performance degradation in CIL.
• To mitigate the aliasing of distributions in incremental learning tasks.

Main Methods:

• Introduced a supervised contrastive learning module with instance- and class-level augmentation during base training.
• Instance-level augmentation involves spatial segmentation at multiple scales to create robust feature representations.
• Class-level augmentation randomly mixes images within mini-batches to promote adaptable decision boundaries.

Main Results:

• The CoLA method enables competitive performance during incremental phases by training only the classifier.
• CoLA+ variant further enhances performance with relaxed data storage constraints.
• Extensive experiments confirm state-of-the-art results across various benchmarks.

Conclusions:

• CoLA effectively balances knowledge acquisition and retention in CIL.
• The proposed augmentation strategies significantly improve performance in incremental learning scenarios.
• CoLA offers a robust solution for continuous learning systems.