My contribution to the “science” vs engineering debate. I think papers should be about explaining reproducible ideas clearly, not about winning datascience competitions. (In the ideal world ;) it should be okay to sacrifice performance for clarity, simplicity and reproducibility. https://t.co/xiBIuspCys
— hardmaru (@hardmaru) July 17, 2018
Ouch: “prestigious institutions had on average 65% higher grant application success rates and 50% larger award sizes, whereas less-prestigious institutions produced 65% more publications and had a 35% higher citation impact per dollar of funding” https://t.co/zRQ9oX0xpR
— Björn Brembs (@brembs) July 16, 2018
I really enjoyed the paper presentation, but the TLDR basically is: If your task cares about absolute positions it makes sense to add a position feature to your input! Do we need a paper for this?
— Denny Britz (@dennybritz) July 16, 2018
So true, unfortunately. In NLP, people have been adding position features to embeddings for a long time. I don’t think anyone has managed to write a whole paper about a one-line feature to make it sound academic. https://t.co/cb9AsRZWMO
— Denny Britz (@dennybritz) July 16, 2018
A+ for this Investigative data-rich journalism. Worthy of best of NateSilver and 538. Careful documentation of research methods as well. Toronto Star stars! https://t.co/JQgBKasCKp via @torontostar
— Edward Tufte (@EdwardTufte) July 15, 2018
Don’t read Data Science Central. The man who runs it wrote this about another data scientist. pic.twitter.com/l1kqknVDpr
— Chris Albon (@chrisalbon) July 15, 2018
Data Scientists: “I don’t have any production code just these Python Jupyter Notebook analyses that inform a multi-billion dollar company’s strategy.” 👏 PRODUCTION 👏 CODE 👏
— Justin Bozonier (@databozo) July 15, 2018
Real risk is incorporating other’s code without understanding or testing it.
— Neil Lawrence (@lawrennd) July 15, 2018
It doesn’t matter who wrote it, bad code is bad code.
But let’s not point at a group and say “don’t write code”, let’s help everyone to write better code.
Computers aren’t smart, they are just dumb really really fast. https://t.co/njPwNIE53A
— bokmann (@bokmann) July 15, 2018
Buggy code is bad science. Poorly tuned benchmarks are bad science. Poorly factored code is bad science (hinders reproducibility, increases chances of a mistake). If your field is all about empirical validation, then your code *is* a large part of your scientific output.
— François Chollet (@fchollet) July 15, 2018
In the context of machine learning research, science and engineering are not distinct concepts. You don't do "science" by thinking very hard about platonic ML concepts and then publishing your thoughts. You do science by engineering systems that test small ideas, iteratively.
— François Chollet (@fchollet) July 15, 2018
OK, as promised here is a small autopsy of a paper I recently came across: https://t.co/uv7Vm0W4YX #Deeplearning #AI https://t.co/6d0GafR84q
— Filip Piekniewski (@filippie509) July 15, 2018
