News for hackers. Stuff that matters.

A comment on Differentiation for Hackers. Parent comment

Re:Differentiation for Hackers by wolfram74 on Saturday June 22, 2019 @12:04PM (#20250285)
Numerical methods can be efficient, but nothing beats an analytic method that can be simply evaluated after derived. Often times you can go from an approximate answer with numerous function evaluations to an exact answer in one simpler function evaluation. Pathological cases not withstanding.
[ Parent | on HN ]
Re:Differentiation for Hackers by adamnemecek on Saturday June 22, 2019 @12:29PM (#20250408)
AD is NOT a numerical method, it's very precise. You can in fact do it on paper as well. https://blog.demofox.org/2014/12/30/dual-numbers-automatic-d...

Analytical derivatives grow exponentially longer with the length of the expression. AD doesn't have this problem.

[ Parent | on HN ]
Re:Differentiation for Hackers by tntn on Saturday June 22, 2019 @01:24PM (#20250670)
Just because automatic differentiation is a different technique than numerical differentiation doesn't mean it isn't a numerical method.

(I probably should not say this, but I will anyway: your comments regarding math and science always give me the distinct impression that you know the name of a lot of things without knowing much about the thing. You are frequently breathless about analog quantum computing, Lie algebras, Chu spaces, some universal concept of duality, but then you fail to respond with substantial arguments to people that try to temper the breathlessness. If you want to convince people of things, you should work on this.)

[ Parent | on HN ]
Re:Differentiation for Hackers by adamnemecek on Saturday June 22, 2019 @03:31PM (#20251279)
> Just because automatic differentiation is a different technique than numerical differentiation doesn't mean it isn't a numerical method.

It does mean that. Pls see the overview paper http://jmlr.org/papers/volume18/17-468/17-468.pdf in particular section "2.1 AD Is Not Numerical Differentiation"

> substantial arguments to people that try to temper the breathlessness

Which arguments in particular?

> our comments regarding math and science always give me the distinct impression that you know the name of a lot of things without knowing much about the thing.

Ok be more specific, which ones do you care about? Maybe I don't always have time to answer everyone.

> some universal concept of duality

It's the concept of adjointness. And yes, it's very universal.

[ Parent | on HN ]
Re:Differentiation for Hackers by tntn on Saturday June 22, 2019 @04:27PM (#20251559)
Huh, interesting. First time I read this comment, it said something like "why should I care what you think?" I guess the edit indicates that you do care?
[ Parent | on HN ]
Re:Differentiation for Hackers by adamnemecek on Saturday June 22, 2019 @04:34PM (#20251586)
No, I edited it because HN has reached out to me many times before that I need to be less of an asshole or I'll get banned.

So no, I don't care about what you think, but I don't want to get banned.

Anyway, nice that you read the first version but still didn't respond to the other questions.

[ Parent | on HN ]