Reproduce “A Probabilistic Framework for Modular Continual Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lazar Valkov, Akash Srivastava, Swarat Chaudhuri, Charles Sutton
PICLE uses probabilistic models to search bounded module compositions for continual learning, combining prior module performance with evidence from new tasks and avoiding training every candidate. On CTrL’s 100-task sequence, it reports 69.65 average accuracy, highest among compared methods; on BELL’s latent-transfer task, forward transfer is 22.28 versus 7.61 for PT-only. Results concern tested benchmarks and path families.
Key insights
- Probabilistic models can effectively estimate the performance of module compositions without full training, significantly reducing computational overhead in modular continual learning.
- Separately optimizing for perceptual/few-shot transfer (via input distribution alignment) and latent transfer (via function similarity using Gaussian Processes) allows for a broader range of knowledge transfer capabilities within a single framework.
- Efficient search strategies, such as greedy search for PT paths and Bayesian Optimization for NT paths, are critical for scaling modular CL to extensive module libraries and long sequences of tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.