Meta-inference and meta-learning

[stuhlmueller]

Inference and optimization should be more "stackable" than they currently are. This issue serves to collect related ideas, some subset of which we may want to tackle jointly.

• Learn Infer/Optimize params by inference/optimization
• Treat learning guides as an inference problem, with priors for guides
• Learn initialization strategies for parameters using infer/optimize
  • This would require us to keep track of randomness during initialization

Some related papers:

• Learning to learn by gradient descent by gradient descent
• Learning to reinforcement learn
• RL^2: Fast Reinforcement Learning via Slow Reinforcement Learning