Research map / Continual learning / Online and replay-based

Trust Region Continual Learning as an Implicit Meta-Learner

NeurIPS 2026 · Continual learning › Online and replay-based · approach: Replay buffer construction and update · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes trust region continual learning, combining generative diffusion replay with a rank-1 Fisher (EWC) penalty, and argues via a one-step MAML-style approximation that it induces emergent fast re-convergence on earlier tasks; evaluated on task-incremental ImageNet-500 diffusion generation and CW10 diffusion-policy control.

Problem

Continual learning faces a tradeoff: EWC-style regularization overconstrains updates when task optima overlap weakly, while replay can retain performance but drift due to imperfect generated replay. The paper also asks why continual models re-converge quickly to earlier tasks.

Approach

Frames replay plus EWC as a Fisher-metric trust region: replay supplies old-task gradients while the rank-1 Fisher penalty anchors updates near old optima. Under local approximations, the combined update is shown to match a one-step MAML meta-gradient, yielding emergent meta-learning behavior without an explicit bilevel objective.

Key results

• On CW10, Trust Region reaches 88.3% average success with forgetting 4.4, versus replay at 85.3% and 8.2 forgetting.
• On ImageNet-500, Trust Region achieves 44.5 average FID and 10.6 forgetting, versus replay at 53.4 and 18.2.
• Re-convergence step counts for Trust Region are often at most 55 steps on ImageNet and 45 steps on CW10 at looser thresholds, while several baselines fail to re-converge.

Surprising findings

• Continual meta-learning baselines (FTML, VR-MCL) underperform finetuning on ImageNet-500 in final FID.
• EWC performs about the same as finetuning on CW10, suggesting a quadratic local surrogate is weak under limited task overlap.

Limitations

• The MAML equivalence relies on local approximations (Hessian approximated by Fisher, rank-1 Fisher) that are assumed rather than verified.
• Meta-learning baselines use first-order approximations at scale, so the comparison may not reflect full bilevel methods.
• Assumes the task-1 optimum lies near a shared optimum; highly heterogeneous task sequences may weaken benefits.

Experimental setup

Datasets: ImageNet-1k (32x32, first 500 classes) Continual-World-10 (Meta-World)

Benchmarks: ImageNet-500 (task-incremental, 32x32) Continual-World-10

Metrics: average FID forgetting average success rate steps to re-converge (task 1)

Baselines: Finetuning EWC Replay FTML VR-MCL

Builds on: elastic weight consolidation model-agnostic meta-learning generative distillation rank-1 empirical Fisher for diffusion models

Closest papers by meaning

• Posterior Meta-Replay for Continual Learning (NeurIPS 2021 · similarity 0.90)
• Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference (ICLR 2019 · similarity 0.89)
• Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning (ICML 2025 · similarity 0.88)
• Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.88)
• Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.86)
• Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.86)