Generalized Variational Continual Learning

• GVCL is a unifying probabilistic framework for continual learning that interpolates between full Bayesian variational inference and local quadratic regularization.
• It employs likelihood tempering and prior scaling to balance plasticity and stability, effectively mitigating catastrophic forgetting across sequential tasks.
• Task-specific FiLM layers and the adaptive AutoVCL mechanism further enhance model performance, calibration, and robustness in nonstationary environments.

Generalized Variational Continual Learning (GVCL) is a unifying probabilistic framework for regularization-based continual learning that interpolates between Bayesian variational inference (as in Variational Continual Learning, VCL) and local quadratic regularization (as in Online Elastic Weight Consolidation, Online EWC). GVCL accomplishes this by introducing likelihood tempering and prior scaling hyperparameters, thereby reconciling fully probabilistic variational approaches and local point-estimated methods. Designed for task sequences requiring adaptation without catastrophic forgetting, GVCL achieves state-of-the-art accuracy and calibration, especially when enhanced with task-specific Feature-wise Linear Modulation (FiLM) layers. Recent extensions such as AutoVCL adaptively schedule the key trade-off hyperparameter using principled heuristics derived from task difficulty and similarity, further improving robustness in nonstationary task environments (Loo et al., 2020, Yang, 2024).

1. Foundational Framework and Objective

GVCL generalizes existing Bayesian continual learning approaches by incorporating likelihood tempering into the standard variational inference objective. For a sequence of $T$ tasks, with each task $t$ associated with dataset $\mathcal{D}_t$, the model maintains at each step a variational posterior $q_{t-1}(\theta)$ over parameters $\theta$, which then serves as a prior when processing new data. The central GVCL loss is the $\beta$-tempered Evidence Lower Bound (ELBO):

$$\mathcal{L}(\theta) = \sum_{t=1}^T \Biggl[ \mathbb{E}_{\theta\sim q_{t-1}(\theta)}\left[ \log p(\mathcal{D}_t\,|\,\theta) \right] - \lambda\,\mathrm{KL}\bigl(q_{t-1}(\theta)\,\|\,p(\theta)\bigr) \Biggr].$$

The regularization weight $\lambda$ (or $\beta$, as used interchangeably in the literature) controls the trade-off between plasticity (incorporating new knowledge from $\mathcal{D}_t$) and stability (preserving information from prior tasks via $t$0). This trade-off is critical for mitigating catastrophic forgetting.

When Gaussian variational posteriors are assumed, decreasing $t$1 (or $t$2) emphasizes local quadratic regularization as in Online EWC, whereas setting $t$3 (or $t$4) recovers full VCL (Loo et al., 2020).

2. Theoretical Interpolation between VCL and Online EWC

GVCL unifies VCL and Online EWC by explicit control of regularization through $t$5 and prior scaling factor $t$6. For task $t$7, the likelihood-tempered objective is:

$t$8

• For $t$9, this coincides with standard VCL: variational inference with full Bayesian uncertainty propagation.
• As $\mathcal{D}_t$0, the variational posterior collapses to a point estimate at $\mathcal{D}_t$1 and the objective approaches Online EWC: a Laplace approximation with Fisher penalty on parameter means accumulated over previous tasks.

Moreover, by introducing prior tempering via $\mathcal{D}_t$2, GVCL exactly recovers EWC's quadratic penalty strength independent of the variational family. This positions GVCL as a spectrum, interpolating smoothly between global probabilistic learning and local, curvature-based regularization (Loo et al., 2020).

3. Algorithmic Formulation and Implementation

The high-level pseudocode of GVCL proceeds as follows (with $\mathcal{D}_t$3, $\mathcal{D}_t$4 denoting mean and covariance parameters):

1. Initialize $\mathcal{D}_t$5.
2. For each task $\mathcal{D}_t$6:
   • Set the prior to $\mathcal{D}_t$7.
   • Compute the tempered KL precision:

   $\mathcal{D}_t$8

   where $\mathcal{D}_t$9 approximates the negative Hessian of expected log-likelihood. - Update $q_{t-1}(\theta)$0 by maximizing

   $q_{t-1}(\theta)$1

• Estimate $q_{t-1}(\theta)$2.

1. Return $q_{t-1}(\theta)$3.

Typical choices are $q_{t-1}(\theta)$4 and $q_{t-1}(\theta)$5, mirroring best practices from EWC (Loo et al., 2020).

4. Mitigation of Variational Over-Pruning via Task-Specific FiLM Layers

A distinct pathology of mean-field VI in deep continual learning contexts is “over-pruning”: many hidden units’ weights collapse to the prior, effectively deactivating units irreversibly. FiLM (Feature-wise Linear Modulation) layers address this by interposing, after each hidden layer activation $q_{t-1}(\theta)$6, a task-specific affine map:

$q_{t-1}(\theta)$7

where $q_{t-1}(\theta)$8 are point estimates unique to each task. In GVCL+FiLM, only shared weights $q_{t-1}(\theta)$9 are regularized by VI, while FiLM parameters are learned by MAP. This scheme allows units “pruned” under one task (e.g., via $\theta$0) to be repurposed in subsequent tasks, thus preserving model capacity and avoiding irreversible capacity loss (Loo et al., 2020).

FiLM layers provide maximal benefit when task identities are explicit. When not available, alternative approaches such as amortized FiLM or learned task conditioners are applicable.

5. Adaptive Scheduling: AutoVCL and Task-Heuristic Modelling

Fixed-choice regularization weights $\theta$1 or $\theta$2 are suboptimal for heterogeneous or non-stationary task streams. AutoVCL extends GVCL by automatically adjusting $\theta$3 per task using measurable task heuristics:

• Task Difficulty $\theta$4: Computed by training a randomly initialized network on a small subset of current task data, normalizing its accuracy relative to random guessing.
• Task Similarity $\theta$5: Measured as the normalized accuracy difference between the prior model (posterior mean) and random guessing on the current subset.

Given the sequence of observed difficulties, the average difficulty gap $\theta$6 and historical maximum difficulty are calculated. AutoVCL sets

$\theta$7

with $\theta$8 a global scaling parameter (commonly $\theta$9). This provides higher $\beta$0 (increased stability) for easier or similar tasks, and lower $\beta$1 (increased plasticity) for difficult or novel tasks (Yang, 2024).

Empirical evaluations show that AutoVCL dynamically tracks optimal regularization, eliminating the need for manual hyperparameter tuning and yielding consistent performance improvements.

6. Empirical Results: Accuracy, Forgetting, and Calibration

Performance of GVCL and its extensions has been assessed across benchmarks including Split MNIST, Permuted MNIST, Split CIFAR-10, and mixed-task pipelines. The addition of FiLM layers and the auto-scheduled regularization of AutoVCL provide:

• Final-task average accuracy for GVCL+FiLM: up to 98.6% on Split-MNIST and 80.0% on Split-CIFAR.
• State-of-the-art results on small-data (e.g., “CHASY” benchmark): substantial gains over HAT, with accuracy improvements in both easy and hard regimes.
• High robustness to catastrophic forgetting (backward transfer near zero), strong forward transfer, and superior expected calibration error compared to MAP-based methods.
• AutoVCL outperforms all fixed-$\beta$2 baselines, with final-task average accuracy on Split MNIST ~97.2% vs. the best fixed baseline at ~96.7%; on Permuted MNIST, ~90.9% vs. ~89.9%; and on mixed MNIST/CIFAR, ~84.9% vs. ~83.7% (Yang, 2024, Loo et al., 2020).

7. Theoretical and Practical Significance

GVCL provides a principled solution for continual Bayesian learning, encapsulating both uncertainty propagation and task-specific regularization via adjustable hyperparameters. The ability to tune the local-vs-global curvature scale with $\beta$3 and control prior strength with $\beta$4 allows adaptation to heterogeneous data regimes. The architecture is computationally efficient when employing diagonal or K-FAC Hessian approximations, with scalability on par with Online EWC. Introduction of FiLM layers is crucial when task identities are known, mitigating over-pruning and maintaining reusable capacity throughout the task sequence. AutoVCL further automates hyperparameter selection, relying only on easily-computable task statistics without retraining or replay buffers.

A plausible implication is that such adaptivity in regularization, coupled with task-modular architectures, may become essential as continual learning grows to embrace more heterogeneous, real-world data streams beyond carefully controlled single-domain benchmarks (Loo et al., 2020, Yang, 2024).