Adaptive Variational Continual Learning via Task-Heuristic Modelling

TL;DR

AutoVCL extends GVCL by introducing a self-adjusted $\beta$-ELBO that adapts to each task through difficulty $d_t$ and similarity $s_t$ metrics. It defines $\beta_t = \exp\left(\lambda\left(\max(d_1,...,d_{t-1}) - \frac{d_t}{1+\delta_d(t-1)} + s_t\right)\right)$ to balance reconstruction versus regularization dynamically, and uses a multi-head architecture with shared parameters updated under $\beta$-ELBO. Across Split MNIST, Permuted MNIST, and mixed MNIST-CIFAR tasks, AutoVCL matches or outperforms fixed-$\beta$ GVCL settings, particularly in later tasks where adaptivity helps preserve prior knowledge while accommodating new tasks. This task-heuristics approach offers a practical, hyperparameter-free pathway to robust continual learning in varied task sequences, with avenues for integrating replay scheduling and smarter difficulty estimation in future work.

Abstract

Variational continual learning (VCL) is a turn-key learning algorithm that has state-of-the-art performance among the best continual learning models. In our work, we explore an extension of the generalized variational continual learning (GVCL) model, named AutoVCL, which combines task heuristics for informed learning and model optimization. We demonstrate that our model outperforms the standard GVCL with fixed hyperparameters, benefiting from the automatic adjustment of the hyperparameter based on the difficulty and similarity of the incoming task compared to the previous tasks.

Figures (2)

• Figure 1: Comparison between AutoVCL and fixed $\beta$ values in three experiments. Split MNIST with Custom Targets, Permuted MNIST, and Mixed Dataset of Split-CIFAR-10 and Split-MNIST respectively represent sequences of similar tasks, different tasks, and alternating difficulty tasks.
• Figure 2: Comparison between tasks 0/1 and 2/3 in the split MNIST experiment. Split 2/3 appears to be a harder task than split 0/1 as its prediction accuracy drops quickly.