Core Concepts
The author introduces a paradigm called Adaptive Discovering and Merging (ADM) to address the challenges of discovering novel classes adaptively while mitigating catastrophic forgetting. The approach involves decoupling representation learning, using Triple Comparison (TC) and Probability Regularization (PR), and proposing Adaptive Model Merging (AMM).
Abstract
The content discusses the challenges of lifelong learning in discovering novel classes from unlabelled data while avoiding catastrophic forgetting. It introduces ADM as a solution that combines adaptive class discovery with model merging techniques to enhance performance. The paper details the methods used, such as TC, PR, AFF, and AMM, along with experimental results showcasing the effectiveness of the proposed approach.
The study compares various state-of-the-art methods in one-step class-iNCD scenarios and demonstrates significant improvements with ADM. Additionally, it evaluates the effectiveness of AMM in class-incremental learning across different stages. The role of gated units in reducing interference between old and new categories is highlighted, showing promising results.
Stats
FRoST achieves 77.5% accuracy for old classes.
FRoST+AFF improves old class accuracy to 77.3%.
FRoST+AMM enhances old class accuracy to 76.8%.
Quotes
"ADM significantly outperforms existing class-incremental Novel Class Discovery approaches."
"Our AMM benefits the class-incremental Learning task by alleviating catastrophic forgetting."