Research map / Continual learning / Online and replay-based
A Probabilistic Framework for Modular Continual Learning
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 PICLE, a modular continual learning framework that uses probabilistic models (a low-rank Gaussian input-distribution model for perceptual transfer and a Gaussian process over module-suffix functions for latent transfer) to cheaply search module compositions, and introduces the BELL compositional benchmark extending CTrL.
Problem
Continual learning with modular networks requires searching a combinatorially large space of module compositions, and evaluating each composition needs costly training, so existing modular CL methods do not scale while achieving perceptual, few-shot and latent transfer.
Approach
Approximate each composition's fitness with a probabilistic model (posterior over pre-trained modules given inputs for perceptual/few-shot paths, and a GP over pre-trained suffix functions for latent paths), so the search trains only a constant number of networks per problem.
Key results
- On BELL short sequences, PICLE achieves average final accuracy 81.13 on the perceptual/few-shot suite versus 79.89 for MNTDP-D, with few-shot transfer Tr^-1 of 46.07 vs 11.42 for MNTDP-D.
- On CTrL S^long (100 problems), PICLE reaches 69.65 average accuracy vs 65.64 for MNTDP-D and 64.49 for LMC.
- On latent transfer sequences in BELL, PICLE achieves 22.97 average Tr^-1 vs 10.8 for MNTDP-D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine perceptual-transfer and latent-transfer search (ablation) | PICLE · BELL latent transfer sequences | average forward transfer: 21.27 -> 22.97 | Helped |
| combine perceptual-transfer and latent-transfer search (ablation) | PICLE · BELL latent transfer sequences | average accuracy: 92.06 -> 92.38 | Helped |
| add latent-transfer search to PT-only (ablation) | PICLE · BELL S-in | forward transfer: 7.61 -> 22.28 | Helped |
| add latent-transfer search to PT-only (ablation) | PICLE · BELL S-sp | forward transfer: 23.65 -> 23.65 | No clear effect |
| add module-accuracy prior for PT paths | PICLE · BELL S-out** | forward transfer: +10.33 | Helped |
| use PICLE rather than MNTDP-D | BELL S-few · BELL S-few | forward transfer: +34.65 | Helped |
| use PICLE rather than PT-only (ablation) | BELL S-in · BELL S-in | forward transfer: +14.67 | Helped |
| use PICLE rather than standalone training | CTrL S-long · CTrL S-long (100 problems) | average accuracy: 55.04 -> 69.65 | Helped |
| use PICLE rather than MNTDP-D | CTrL S-long · CTrL S-long (100 problems) | average accuracy: 65.64 -> 69.65 | Helped |
| use PICLE rather than MNTDP-D | BELL compositional short sequences · BELL compositional short sequences | average forward transfer: 10.8 -> 17.22 | Helped |
| use PICLE rather than MNTDP-D | CTrL short sequences · CTrL short sequences | average accuracy: 65.58 -> 66.00 | Helped |
Surprising findings
- Modular methods all fail to achieve backward transfer because pre-trained modules are frozen.
- LMC performed worse than standalone on the BELL sequences despite adjustment efforts.
Limitations
- PT prior relies on accuracy, hard to apply to tasks like regression or differing task types.
- NT prior restricts to suffixes derived from previous solutions, precluding novel suffixes.
- Does not search over paths combining initial and final pre-trained layers.
Experimental setup
Datasets: CTrL CIFAR100 BELL
Benchmarks: BELL CTrL
Metrics: average accuracy forward transfer on last problem (Tr^-1) average forgetting
Baselines: SA O-EWC ER RS HOUDINI MNTDP-D LMC
Builds on: CTrL benchmark HOUDINI MNTDP-D LMC PathNet
Closest papers by meaning
- MoRe: Modular Representations for Principled Continual Representation Learning on Sequential Data (NeurIPS 2026 · similarity 0.88)
- TinySubNets: An Efficient and Low Capacity Continual Learning Strategy (AAAI 2025 · similarity 0.87)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.87)
- Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning (NeurIPS 2025 · similarity 0.87)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.87)
- Optimal Task Order for Continual Learning of Multiple Tasks (ICML 2025 · similarity 0.86)
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.