As a sr data scientist who also does a lot of eng – “uh let me get back to you on that (after I do some research)” is still a VERY common answer to a lot of questions I get. I STILL use the same process with new problems: fail a lot, learn, iterate, break up into smaller problems https://t.co/5DRrPeHqrA
— Mikhail Popov (@bearloga) October 9, 2020
Programming: 10% writing code. 90% figuring out why it doesn’t work
— Ben Hamner (@benhamner) October 9, 2020
Analyzing data and ML: 1% writing code. 9% figuring out why code doesn’t work. 90% figuring out what’s wrong with the data
Don't cook up use cases for #AI when simple human input solved the problem. At best it is a poor substitute for human. At worst you are propagating #AI #bias ans amplifying and automating it https://t.co/uU6MQvbrRT
— Prof. Anima Anandkumar (@AnimaAnandkumar) October 1, 2020
2. You only have a few ML models in production
— Chip Huyen (@chipro) September 29, 2020
Booking, eBay have 100s models in prod. Google has 10000s. An app has multiple features, each might have one or multiple models for different data slices.
You can also serve combos of several models outputs like an ensemble.
(3/6)
When talking to people who haven’t deployed ML models, I keep hearing a lot of misperceptions about ML models in production. Here are a few of them.
— Chip Huyen (@chipro) September 29, 2020
(1/6)
Saying that bias in AI applications is "just because of the datasets" is like saying the 2008 crisis was "just because of subprime mortgages".
— François Chollet (@fchollet) September 20, 2020
Technically, it's true. But it's singling out the last link in the causality chain while ignoring the entire system around it.
I've received so much criticism for not incorporating reward info into world model / representations used by RL agents
— hardmaru (@hardmaru) September 18, 2020
But the way I see it, rewards are so overvalued…
See this new paper, “Decoupling Representation Learning from Reinforcement Learning”https://t.co/ADKKpOJLGI https://t.co/APTEJRMmOs
"Statistical significance in deep learning papers?" (https://t.co/RB39xlkjrh) -- why is this not a thing in practice? My thoughts: (a) test sets are usually large enough; (b) reviewers trad. didn't ask for it (c) people don't like pitfalls (independence violations). Others?
— Sebastian Raschka (@rasbt) September 17, 2020
I think the biggest difference is that ML people are trying to build software systems, whereas statisticians seek to support scientific inquiry. This is why ML folks don't typically care about estimation or statistical inference
— Thomas G. Dietterich (@tdietterich) September 9, 2020
Change my mind: the biggest difference between statistics and machine learning is that statisticians think a lot about sampling and people in ML largely...don't.
— Kristian Lum (@KLdivergence) September 9, 2020
Some great advice from @visnut https://t.co/P3jp1zxvYf
— Rob J Hyndman (@robjhyndman) August 28, 2020
You understand a thing if you can recreate it -- not in the weak sense of being able to *copy* it, but in the strong sense of being able to generate deliberate variations of it. Including improvements.
— François Chollet (@fchollet) August 25, 2020
You understand a thing if you can recreate it *better*.
