December 06, 2017

On machine learning and programming languages

Julia Blog, Reddit Discussion, SysML Paper, By Mike Innes (Julia Computing), David Barber (UCL), Tim Besard (UGent), James Bradbury (Salesforce Research), Valentin Churavy (MIT), Simon Danisch (MIT), Alan Edelman (MIT), Stefan Karpinski (Julia Computing), Jon Malmaud (MIT), Jarrett Revels (MIT), Viral Shah (Julia Computing), Pontus Stenetorp (UCL) and Deniz Yuret (Koç University).

Summary: Any sufficiently complicated machine learning system contains an ad-hoc, informally-specified, bug-ridden, slow implementation of half of a programming language.

1 comment:

Emre Sevinc said...

Maybe I'm misinterpreting it, but I started to think that the motivation and the direction mentioned here share some points with what Conal Elliott is trying to achieve in The simple essence of automatic differentiation - (Differentiable functional programming made easy).

I find it exciting to observe the trends and where these efforts will lead to in 5 to 10 years: shall we have much better data science and machine learning systems? (Without huge commercial support from the usual suspects such as Google, MS, Amazon, Facebook, etc.)