News for hackers. Stuff that matters.

A comment on Data scientists need to learn about significant digits.

Re:Data scientists need to learn about significant digits by tobinfricke on Monday December 25, 2023 @05:29PM (#38766828)
Omg yes. My coworkers often specify parameters (the results of what we would have called "regression" or "fitting" but what is now called "machine learning") to like 14 decimal places when we only know them to 20%. It would have yielded a failing grade in my high school chemistry class.

They are like "Chesterton's digits" when they appear unexplained in parameter files. Do we really know this number or is it just a wild-assed guess?

[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by bigbillheck on Monday December 25, 2023 @05:50PM (#38767045)
> Do we really know this number or is it just a wild-assed guess?

Does it matter? You have 23 bits of mantissa to fill one way or the other, and if you don't know the lower, say, 12, one pattern of bits is as good as any other pattern.

[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by diarrhea on Monday December 25, 2023 @06:11PM (#38767226)
Precision beyond what can be justified by the uncertainty of the original input is wrong, not just as-good-as.
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by bigbillheck on Monday December 25, 2023 @07:49PM (#38767870)
In terms of actual numbers in the physical world, sure, but the usual kind of floating point numbers don't have a concept of precision, and you're getting all 23 bits whether you want them or not.

Humans know the difference between "1.0" and "1.0000" but they're the same to the computer.

[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by rcxdude on Tuesday December 26, 2023 @06:46AM (#38770956)
rounding, especially in decimal, isn't actually any better. If you want to specify precision, specify it explicitly.
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by tobinfricke on Monday December 25, 2023 @06:14PM (#38767242)
That would be fine if the number were accompanied by a confidence interval.
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by pessimizer on Monday December 25, 2023 @06:28PM (#38767335)
The reason you get rid of those extra digits is because it will always be tempting to find meaning in them.
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by tobinfricke on Monday December 25, 2023 @06:58PM (#38767527)
Great answer!
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by dickersnoodle on Monday December 25, 2023 @05:43PM (#38766981)
puts on grumpy old man hat Yeah, back in my day you had to learn about significant digits because we still used slide rules and the calculators didn't have as many functions and we liked it! removes grumpy old man hat
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by tobinfricke on Monday December 25, 2023 @06:13PM (#38767237)
That's not the reason sigfigs are important.
[ Parent | on HN ]
Re:Data scientists need to learn about significant digits by dickersnoodle on Saturday January 06, 2024 @10:41AM (#38892302)
Oh, I'm aware. On the way to getting a bachelor's in Chemistry I was drilled on the importance of sigfigs as a way of preventing data contamination by artificial precision (e.g. trying to weigh 16.0000000000 lbs with a bathroom scale). It just seemed funnier this way...
[ Parent | on HN ]