Open questions to answer with real data

[fritzo]

• What should be the default thompson_temperature? 1.0?
• Will inference="svi" work, or will we need the more expensive inference="mcmc"?
• Is our (0,1)-valued likelihood sufficient for real cases?
• Are our default hyperparameters that control batch structure reasonable?

cc @andrenguyen