Concetti Chiave
Enhancing forward compatibility in class incremental learning by increasing representation rank and feature richness.
Sintesi
Continual learning challenges conventional static datasets.
Class Incremental Learning (CIL) focuses on adaptive learning for new classes.
Backward compatible methods address catastrophic forgetting.
Forward compatible approaches aim to enhance subsequent task training.
RFR method increases effective rank for richer features and forward compatibility.
Theoretical connection between effective rank and Shannon entropy proven.
Empirical investigations show the effectiveness of RFR in enhancing novel-task performance and mitigating forgetting.
Extensive experiments validate the efficacy of RFR across eleven CIL methods, improving average incremental accuracy.
Statistiche
RFR achieves enhancements of 2.50%, 2.63%, and 2.59% for split sizes of 10, 5, and 2 respectively.
Effective rank is defined as exp(-∑𝑖=1 𝜆𝑖log 𝜆𝑖).
Empirical results show an average reduction in weight distances with RFR across all sessions.
Citazioni
"In this study, we introduce an effective-Rank based Feature Richness enhancement (RFR) method."
"Our method can effectively increase the representation rank."
"Our results demonstrate the effectiveness of our approach in enhancing novel-task performance."