Differentiable programming for gradient-based machine learning

I think the ownership system, when fully elaborated, may effectively give us noescape results. A yield'ed borrow of a non-copyable type is, for example, non-escaping. The problem is that closures are copyable. I haven't read up on the plans for making closures with mutable captures threadsafe, but one way would be to make them move-only. I bring this all up just to point out that there may be some interesting design synergies here that should be explored.

5 Likes