Updated ChainRulesCore.rrule

[alfredjmduncan]

Replaces the pullback for axes argument with NoTangent().

[jonniedie]

Hey, so my knowledge of ChainRules is admittedly pretty low. Would you mind explaining why we wouldn't need the axes in the pullback? Or I guess what we get from using NoTangent here? Is this just a performance thing so we can eliminate what's likely to be a dead branch in the pullback? Or is this producing something incorrect?

[jonniedie]

Or is the idea that differentiating through Axis indices is not going to be meaningful anyway?

[alfredjmduncan]

Or is the idea that differentiating through Axis indices is not going to be meaningful anyway?

That's right, apologies for the short PR description. As the Axis is discrete, no derivative exists for it. My original motivation was that I was running into an error using ComponentArrays with AdvancedVI.jl, where the existence of the pullback for the Axis was causing a problem. But my understanding is that the pullback for the axes argument should in theory be a NoTangent(). Tagging @oxinabox and @ChrisRackauckas as we had a brief exchange on Slack.

[jonniedie]

Okay cool. That makes sense. I’ll go ahead and merge it.

Thanks for the pr!