https://developer.apple.com/videos/play/wwdc2018/609/

I found the presentation very clear and the api looks very easy to understand and seems like it would be easy to use.

I’m curious about the design decision not to automatically synthesize the backprogation half of the graph. It makes for a very nice symmetry to explicitly define the gradient pass it as they’ve done but I’m curious if there are practical reasons a practitioner might need/want access to the explicitly defined back propogation graph?

Is the difference in modeling in MPS vs say tensorflow mainly a result of the fact that MPS is a “lower level” api than tensorflow? Are there maybe some useful manual tweaks a practitioner might make to the backprop graph vs something synthesized automatically — or useful observations of the gradient layer outputs that a practitioner might be interested in?