ConceptAI & ML
Distribution shift
A change in the data distribution between the training phase and the inference/test phase.
Know first
Named as a prerequisite by the lectures that define this. The number is how many say so. A grey one has no page yet.
Probability distribution3Supervised learning2Imitation learning1Markov decision processes1Model correction1Offline reinforcement learning1Reinforcement learning1Training and test data sets1
Opens up
Concepts that name this one as a prerequisite.
Lecturers
5Each in their own terms, from the easiest level.
Sasha Rush
Actual mistakes deviating from training examples during reinforcement learning for model correction.Sergey Levine
The problem where a model is trained on one particular distribution and then needs to perform well under a different distribution.Andrea Zanette
The mismatch between the state-action distribution present in a historical dataset and the distribution induced by the target optimal policy.Russ Tedrake
Distribution shift occurs when a closed-loop policy drifts away from the training data distribution, leading to compounding errors.
Lectures that assume it
3Connections
17The lectures that define it or assume it, and the concepts next to it.
Concepts9
- Probability distribution
- Supervised learning
- Imitation learning
- Markov decision processes
- AoTiL (accuracy-on-the-line)
- Covariate shift
Show 3 moreShow fewer
Lectures8
- Speculations on Test-Time Scaling for Large Language Models (o1)
- Offline Reinforcement Learning
- Generalization: Out-of-Distribution (OOD)
- Towards a Statistical Foundation for Reinforcement Learning
- Imitation Learning, Foundation Models, and Course Wrap-up
- Transfer Learning: Models