How Do You Decide What Is in or Out of Distribution?

It's really tricky to define those things in reinforcement learning because you have a set of tasks or distribution of tasks that are sort of pan designed and often are not actually a distribution. And one of the things that we found, even in like multitask learning settings, is how different two tasks are, two distributions are. It depends on the data itself, but can also depend on the model and how it learned. We're building tools for trying to evaluate that.