Keynote: Automatic Differentiation for Dummies

[@rrr00bb1]

this is by far the most vexing explanation ever.

[@gabrielmccartney7975]

Hello! Can we use dual numbers for integration?

[@ElizaberthUndEugen]

Why `d f : S -> (S -o T)` and not `d f : (S -> T) -> (S -o T)`? After all `d` takes a function and returns a linear map. The notation says it takes an `S` and returns a linear map. huh? Edit: Ah, because the linear map is a different one depending on where on the domain of `f` the derivative is evaluated. But how does that match the definition of the derivative as a limit?

[@kanchankumar1519]

'df' itself is the function being defined here, (d is an operator and isn't defined here). Also, the notion of limit leads us to a linear function as derivative. In 1d for example, the derivative is the tangent line, which is simply a linear function. In 2-d, it would be a plane -- again a linear map

[@sapito169]

this is not for dummies

[@JSimba94]

I felt like a dummy watching. But a great watch regardless!

[@guzh]

@@JSimba94 confusing lecture. NG.

[@haps3000]

Also, congrats to the audience for supressing their inner rage towards the guy that created Haskell.

[@undisclosedmusic4969]

After 1 hour and 5 minutes of pretentious programming language and notation BS, turns out this magical implementation is still just plain, old reverse-mode AD. Still gotta define the local gradients, still gotta backpropagate…

[@vtrandal]

Thank you! Seriously, thank you for sorting this mess for me. To make matters worse the speaker is chummy and on a first name basis with Conal Elliot.