Assignment #12
Provide a short discussion of each of the assigned papers (listed under Course Materials). Below are some questions to think about.
MAML
Questions
- Meta-learning learns a learning algorithm. In MAML, which aspects of the learning algorithm are being meta-learned? Which are hardcoded?
- In Figure 3, MAML (green curve) improves very rapidly after just one step of gradient descent, but for additional steps the improvements are no better than the pretrained baseline (blue curve). Explain this behavior. (Similar behavior can be observed in Figs 4 and 5.)
RL^2
Questions
- Which aspects of the learning algorithm are being meta-learned? Which are hardcoded?
- The "fast RL algorithm" is optimized to learn from a fixed number of episodes, e.g., 2 episodes for the maze navigation experiments. What limitations might this lead to? Often, we would like learners that can learn well from any amount of experience. Can you think of ways to modify RL^2 such that it could learn well from any amount of experience?
Modular Meta-Learning
Questions
- Which aspects of the learning algorithm are being meta-learned? Which are hardcoded?
- What's an advantage of meta-learned modules over the approach in MAML? What's a disadvantage?
General questions:
- Describe a case where meta-learning will be detrimental (i.e. perform worse than learning from scratch).
- Meta-learning is learning to learn. This idea can be applied recursively: we can learn to learn to learn. Can you think of an example (in humans or machines) of learning to learn to learn? Where might this be useful?
- Describe the relationship between MAML, RL^2, and Modular Meta-Learning. Which is most expressive? Can one be seen as a special case of another?
Upload a single PDF file through Stellar by Mar 21 at 10 am.