Core Concepts
Effective-Rank based Feature Richness enhancement (RFR) method improves forward compatibility in class incremental learning by increasing representation rank.
Abstract
Continual learning challenges conventional static datasets.
Class Incremental Learning (CIL) focuses on adaptive learning for new classes.
RFR method enhances feature richness by increasing effective rank during the base session.
Theoretical connection between effective rank and Shannon entropy is established.
Extensive experiments validate RFR's effectiveness in enhancing novel-task performance and mitigating catastrophic forgetting.
RFR consistently improves performance across eleven existing methods.
RFR shows promise in non-exemplar-based approaches, larger-scale datasets, and models.
Stats
Effective rank is a continuous-value extension of algebraic rank.
Representation rank can serve as an indicator of the quantity of encoded features.
Effective-Rank based Feature Richness enhancement (RFR) method increases representation rank during the base session.
Quotes
"Representation rank can serve as a crucial indicator of the quantity of encoded features."
"RFR achieves two distinct methodological objectives solely through a forward compatible approach."