Mike J Innes

I haven't followed this issue carefully but (1) yes, Zygote supports structs well and (2) it'd be nice not to have to load `DistributionsAD` on top of `Distributions` to get...

Lyndon mentioned it, but just linking [ZygoteRules](https://github.com/FluxML/ZygoteRules.jl) explicitly. RE using Zygote in the wild: the marker for that is really going to be when we release Flux + Zygote; once...

You could try explicitly dropping gradients of globals to see if that's the issue.

The `rrule`/`frule` overload also starts to get more complex when you're passing contexts around, which is necessary to handle mutation. That's a good reason to have both a rule-making macro...

I'm really sceptical that a sensible design can be had in which we have both composite and abstractarray gradients. Will's concern in https://github.com/FluxML/Zygote.jl/pull/445 was that using a FillArray as the...

> At some point you'll have to accumulate the differential for `x[1]` and `x[2]`, whose sum is clearly not also a 1-hot array. Also, while I appreciate that this is...

It depends on what the exact proposal is; the discussion on the Zygote issue indicated a mixing of array and composite adjoints (i.e. using composites for efficiency where possible but...

Assuming you buy the `xs[i] ≡ mean(xs)` argument, then `xs[i]` is also equivalent to any weighted mean of `xs` (where the weights sum to one). With random weights you get...

The core point is really that switching to `Composite` gradients changes the gradients we get for `W` and `x`, in some cases, compared to the current situation; this is what...

I have some PRs open for Atom which are aimed at fixing this: https://github.com/atom/atom/pull/11864 & https://github.com/atom/atom/pull/11852.