When incrementally trained for new tasks, deep continual learning models typically suffer from catastrophic forgetting, where parameter updates optimized for new tasks deteriorate performance on previously mastered tasks. This paper proposes a comprehensive replay-based approach built upon the OnPro model to address this fundamental challenge through three synergistic contributions. First, we employ Equilibrium Training with Prototypical Feedback to enhance discriminative feature quality through adaptive prototype evolution. Second, we introduce a specialized multi-head architecture incorporating (i) within-task prediction (WP) heads for task-specific classification and (ii) out-of-distribution (OOD) heads for inter-task class separation. Third, we develop an Energy-based Latent aligner (ELI) that learns energy manifolds to minimize representational drift, assigning higher energy to current task representations and lower energy to previous task features. Our method demonstrates substantial improvements on medical imaging (CCH5000) and standard benchmarks (CIFAR-10/100), achieving 27.5\% forgetting reduction on medical datasets and up to 21.4\% accuracy improvement on CIFAR-10 under memory constraints. However, our analysis reveals important trade-offs: computational overhead increases by 82\%, and effectiveness diminishes in single-class scenarios due to insufficient negative sampling diversity for contrastive learning. These findings provide both methodological advances and theoretical insights that guide future continual learning research.
Thi et al. (Fri,) studied this question.