Learning without losing old skills
Learn new tasks while preserving useful old knowledge.
What you’ll explore
- Verifiers
- Curricula
- Forgetting
- Meta-learning
Your lessons
0 / 9 completeBuild a trustworthy practice loop
Generate examples, filter them with a verifier, train a candidate, and protect earlier skills before promotion.
Learn something new without losing the old
Train a shared neural network on two tasks, measure forgetting, and compare replay with a parameter-change penalty.
Learn a better way to adapt
Optimize a starting weight for rapid adaptation, then evaluate adapted copies on a task excluded from meta-training.
Choose what experience to retain
Compare reservoir, task-stratified and recent replay under a frequency shift, from the same neural checkpoint.
Change context or change parameters
Give identical support measurements to a context fit, exact-key memory and gradient-based parameter adaptation.
Measure adaptation cost
Compare zero and meta-trained initializations across linear and shifted curved task families using fixed support/update budgets.
Learn reusable temporally extended skills
Train two tabular option policies and logistic termination functions, then compare staged navigation with a flat learner and fixed hierarchy.
Coordinate agents with limited information
Compare centralized observations, local missing information and explicit communication in a synchronous two-agent grid.
Update without silently changing the contract
Train a real candidate, recompute old/new regression gates and retain both checkpoints for a replayable rollback record.