Q&A: Phase 33.4 CurriculumScheduler — Design Questions

[web3guru888]

Q&A: Phase 33.4 — CurriculumScheduler Design Questions

This thread is for technical questions about the CurriculumScheduler component.

Open design questions

1. Task embedding space: How do we compute meaningful task similarity without training? Options include dataset statistics (mean/variance), Fisher information overlap, or learned task embeddings. Which is most predictive of actual transfer?

2. Optimal ordering complexity: Finding the truly optimal task ordering is NP-hard (TSP variant). Our greedy chain heuristic is O(T²) — is this sufficient, or should we invest in more sophisticated approximations (e.g., beam search, genetic algorithms)?

3. Spaced repetition intervals: The Leitner intervals [1, 3, 7, 15, 30] are borrowed from flashcard learning. Are these appropriate for neural network continual learning, or should intervals be adapted based on measured forgetting curves?

4. Online vs. offline scheduling: If the task pool is not known in advance (tasks arrive streaming), how does the scheduler adapt? Should it maintain a buffer of pending tasks to reorder?

5. Interaction with replay: If replay (Phase 33.3) already revisits old tasks via buffer, does spaced repetition provide additional benefit? When should we use scheduling vs. replay vs. both?

References

• Issue: Phase 33.4 — CurriculumScheduler: Intelligent task sequencing for lifelong learning #694
• Planning: Phase 33 Planning — Continual Learning & Catastrophic Forgetting Prevention #690