Adam Haber

This just happened to me on macOS as well.

I'll try! Will submit a draft PR once I have something working.

@sritchie are the inputs to the metric comparable? For example, if we'll wrap the memoized metric with a function that ensures that `metric x y` is always called with `x>y`......

Well, I tried this and I guess the answer is no? ``` ERROR in () (Numbers.java:259) expected: (= (quote (/ (+ (* -8 G pi (expt (R t) 2) (rho...

What would be the best place to put this functionality (new ns, `complex`, `euclid`, etc)?

I'll try!

Another thing we might want to look into is implementing [checkpointing](https://pytorch.org/docs/stable/checkpoint.html). The motivation here is that a lot of physical computations end up as differential equations, and checkpointing can be...

Interesting, I'll try that, thanks! The problem is that my current precomputed distance matrix is 1000x1000, and I'm not sure if there's any particular way to choose which are the...

Thanks for the link @bellet , seems super interesting! As for the reason for learning a precomputed metric - exactly the two reasons you've mentioned. :-) I am indeed interested...

Indeed, I mean scaling with the number of features. SGD scales well with the number of features _and_ the number of points, but it's not clear how to impose the...