Research map / Continual learning / Online and replay-based
Functional Regularisation for Continual Learning with Gaussian Processes
Michalis K. Titsias, Jonathan Schwarz, Alexander G. de G. Matthews, Razvan Pascanu, Yee Whye Teh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use trace-term optimisation instead of random selection for inducing points (ablation) | FRCL · Permuted-MNIST, 10 points/task | accuracy: 48.2% { $\pm$ 4.0} -> 61.7% { $\pm$ 1.8} | Improved |
| Use trace-term optimisation instead of random selection for inducing points (ablation) | FRCL · Split-MNIST, 2 points/task | accuracy: 59.8% { $\pm$ 8.0} -> 82.0% { $\pm$ 5.0} | Improved |
| Use trace-term optimisation instead of random selection for inducing points (ablation) | FRCL · Omniglot, 1 point/char | test accuracy: 69.74 { $\pm$ 1.1} -> 72.02 { $\pm$ 1.3} | Improved |
| Use trace-term optimisation instead of random selection for inducing points (ablation) | FRCL · Omniglot, 3 points/char | test accuracy: {\bf 81.42} { $\pm$ 1.2} -> 81.01 { $\pm$ 1.1} | Hurt |
| Increase inducing points from 1 to 3 per character (ablation) | FRCL (Trace) · Omniglot | test accuracy: 72.02 { $\pm$ 1.3} -> 81.01 { $\pm$ 1.1} | Improved |
| Increase inducing points from 1 to 3 per character (ablation) | FRCL (Random) · Omniglot | test accuracy: 69.74 { $\pm$ 1.1} -> {\bf 81.42} { $\pm$ 1.2} | Improved |
| Use trace-term rather than random inducing point selection (ablation) | FRCL · Permuted-MNIST, 200 points/task | accuracy: 94.2% { $\pm$ 0.1} -> 94.3% { $\pm$ 0.2} | No clear effect |
| Use function-space posterior summaries and KL regularisation instead of an explicit replay buffer | FRCL versus replay baseline · Permuted-MNIST, 200 points/task | accuracy: 82.3% { $\pm$ 0.3} -> 94.3% { $\pm$ 0.2} | Improved |
| Use function-space posterior summaries and KL regularisation instead of an explicit replay buffer | FRCL versus replay baseline · Split-MNIST, 40 points/task | accuracy: 95.8% { $\pm$ 1.1} -> 97.8% { $\pm$ 0.7} | Improved |
| Use function-space posterior summaries and KL regularisation instead of an explicit replay buffer | FRCL versus replay baseline · Omniglot, 1 point/char | test accuracy: 42.73 { $\pm$ 1.2} -> 72.02 { $\pm$ 1.3} | Improved |
| Use function-space posterior summaries and KL regularisation instead of an explicit replay buffer | FRCL versus replay baseline · Omniglot, 3 points/char | test accuracy: 65.32 { $\pm$ 1.1} -> 81.01 { $\pm$ 1.1} | Improved |
| Use FRCL trace selection rather than VCL with random coreset | Omniglot MLP · Omniglot, 3 points/char | test accuracy: 51.64 { $\pm$ 1.0} -> 53.86 { $\pm$ 2.3} | Improved |
Closest papers by meaning
- Continual Learning via Sequential Function-Space Variational Inference (ICML 2022 · similarity 0.91)
- Continual Deep Learning by Functional Regularisation of Memorable Past (NeurIPS 2020 · similarity 0.91)
- A Statistical Theory of Regularization-Based Continual Learning (ICML 2024 · similarity 0.88)
- Optimal Rates in Continual Linear Regression via Increasing Regularization (NeurIPS 2025 · similarity 0.87)
- Function-space Parameterization of Neural Networks for Sequential Learning (ICLR 2024 · similarity 0.87)
- Learning Representations on the Unit Sphere: Investigating Angular Gaussian and Von Mises-Fisher Distributions for Online Continual Learning (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
