FT-IBCL: Enhancing Imprecise Bayesian for Continual Learning Under Trade-Offs with Fine-Tuning
Quynh-Trang Pham Thi, Tu-Tai Hoang, Van-Toan Phan, Tri-Thanh Nguyen, Quang-Thuy Ha, Thanh Hai Dang · 2024
Continual learning face persistent challenges in balancing stability (retaining knowledge of previously learned tasks) with plasticity (adapting to new tasks). While recent approaches like Imprecise Bayesian Continual Learning (IBCL) have shown promise in managing this trade-off through zero-shot adaptation, they may not fully optimize for task-specific performance. We propose FT-IBCL (Fine-tuned Imprecise Bayesian Continual Learning) model that enhances IBCL by incorporating a supervised fine-tuning phase. FT-IBCL maintains a knowledge base using Finitely Generated Credal Sets (FGCS) and leverages Highest Density Regions (HDR) for zero-shot model generation, while introducing targeted fine-tuning using a small subset (15 %) of task-specific data. Comprehensive experiments on two benchmark datasets, i.e. CIFAR-10 and CIFAR-100, demonstrate that FT-IBCL significantly outperforms the original IBCL and other baseline methods, particularly in peak accuracy (achieving 91.9% on CIFAR-10 and 92.2% on CIFAR-100) and backward transfer metrics. Experimental results indicate that FT-IBCL successfully bridges the gap between zero-shot adaptation and task-specific optimization, offering an efficient and effective approach for continual learning scenarios.