|
Wired has more details on how it actually works and uhquote:The idea was to use prior patterns to infer what a student would have scored in a 2020 not dominated by a deadly pandemic. IB did not disclose details of the methodology but said grades would be calculated based on a student’s assignment scores, predicted grades, and historical IB results from their school. The foundation said grade boundaries were set to reflect the challenges of remote learning during a pandemic. For schools where historical data was lacking, predictions would build on data pooled from other schools instead.
|
# ? Jul 14, 2020 16:34 |
|
|
# ? May 27, 2024 03:16 |
|
I'd say people should sue, but that requires a judge with a reasonable understanding of machine learning, and wow is that a bad bet.
|
# ? Jul 14, 2020 16:36 |
|
I think "it's assigning you grades not based on your own grade, but on what other people have gotten as grades in the past" is fairly easy to understand! and in any case discovery should turn up the model they use and how they calibrated it, which would be uh really interesting
|
# ? Jul 14, 2020 17:10 |
|
There actually is some work on predicting scores for parts of a candidate's bar exam if they get lost. See https://thebarexaminer.org/wp-content/uploads/PDFs/780309_Mroch.pdf for discussion. My guess is that whatever the IB committee did is less thoughtful. I can't find it now, but I remember at some point reading a complaint from a candidate who completely lost their poo poo when their score on a part of an exam was predicted and they failed. It had every combination of bold, underline and italics that there is.
|
# ? Jul 14, 2020 17:18 |
|
Pinterest Mom posted:I think "it's assigning you grades not based on your own grade, but on what other people have gotten as grades in the past" is fairly easy to understand! or they immediately settle and give everyone the equivalent of As to moot the case to avoid discovery
|
# ? Jul 14, 2020 17:53 |
|
that paper is p thoughtful, yeah, thanks. it's just using OLS but honest about it. they're really helped by the fact that the bar is P/F, and using a decision rule like "just pass them if we predict there's a 10% chance they would have passed" is sensible. predicting a score along a continuum is a lot harder, and the way that IB just outright used demographics, whew.
|
# ? Jul 14, 2020 17:55 |
|
https://www.youtube.com/watch?v=aJzmCUvLNrI
|
# ? Jul 17, 2020 15:06 |
|
its so incredibly shameful that the software is caled waifu
|
# ? Jul 17, 2020 15:06 |
|
fart simpson posted:andrew ng came to my office in shenzhen for a meeting and i saw him and he looked at me few years back i went to an andrew ng talk about "the future of ml". he said that current machine learning can do anything a human can do in less than two seconds. so, checking a license, reacting to a curve in a road, reading a bit of text; he said all that can be automated in the next 10 years. (he was lying through his teeth, but whatever.) during the q&a i asked what was gonna happen to all the people whose jobs involve lots of short repetitive tasks. he said, well, we're creating a lot of jobs in the machine learning space.
|
# ? Jul 18, 2020 03:31 |
|
There are several million long distance truckers in the US, and they collectively have a lot of guns and meth. Do we really want to make them unemployable?
|
# ? Jul 18, 2020 04:05 |
|
yea that sure sounds scary but on the other hand, every other industry also had guns and drugs & yet they still became unemployed
|
# ? Jul 18, 2020 04:38 |
|
did he at least look ashamed when he said that
|
# ? Jul 18, 2020 06:47 |
|
https://twitter.com/isosteph/status/1287617727943720961
|
# ? Jul 27, 2020 13:58 |
|
|
# ? Jul 27, 2020 15:24 |
|
"im not an algorithm! im not an algorithm!!", i continue to insist as i slowly shrink and transform into a multilayer perceptron
|
# ? Jul 27, 2020 16:50 |
|
https://twitter.com/themarkup/status/1286270083791118336 https://twitter.com/JuliaAngwin/status/1287795930578722817
|
# ? Jul 27, 2020 20:49 |
|
jfc
|
# ? Jul 27, 2020 20:53 |
|
ultrafilter posted:https://twitter.com/themarkup/status/1286270083791118336 amazing
|
# ? Jul 27, 2020 20:58 |
|
I developed a pathon machine learning discord bot that can detect anime avatars with a decent success rate. It then gives me a list of people i can ban So far its working out e: i needed an nvidia library for that and I had to agree to not use it to develop softeare to be used for evil and bullying people Michaellaneous fucked around with this message at 17:00 on Jul 28, 2020 |
# ? Jul 28, 2020 16:38 |
|
https://twitter.com/tenuous/status/1288174286042288133
|
# ? Jul 28, 2020 19:55 |
|
hooooly gently caress, even for the abysmal standards or acceptability in machine learning startups that is a horrible idea
|
# ? Jul 28, 2020 20:01 |
|
hm (btw i assume everyone in yospos is worldly enough to know this, but just in case, don't put your real name in there)
|
# ? Jul 28, 2020 20:04 |
|
Might need a bigger training set, or a different algorithm, or a fundamental rethinking.
|
# ? Jul 28, 2020 20:04 |
|
Like, ok: so let's play arguendo. Set aside the fact it doesn't seem to loving work (assume that could be fixed): what's the value proposition? What is the path to revenue here?
|
# ? Jul 28, 2020 20:07 |
|
quote:Genderify primarily focuses on the B2B business model, and our approach for classifying names, usernames, or emails as male or female is free of any gender discrimination quote:And though in many cultures, gender may generally be viewed as binary, recognizing genders mostly as male and female, there are many more gender identities people use to define themselves and express their gender. As our AI model’s decisions are based on already existing binary name/gender databases, our Product team is actively looking into ways of improving the experience for transgender and non-binary visitors. For example, the team is working on separating the concepts of name/username/email from gender identity. In that sense, we are pioneers in developing new approaches to provide non-binary results, and we do very much rely on community feedback to make it work. Through a special feedback system, we are going to improve our gender detection algorithms for the LGBTQ+ community to make sure we can break the old school bias association between name to gender. Dunno
|
# ? Jul 28, 2020 20:10 |
|
Schadenboner posted:What is the path to revenue here? Step 1: Machine Learning Step 2: ??? Step 3: Profit
|
# ? Jul 28, 2020 20:11 |
|
ultrafilter posted:Might need a bigger training set, or a different algorithm, or a fundamental rethinking. this is one where no amount of technology, thinking, and ethical awareness fixes the fundamental problem: hearing someones name and having a guess at their gender is socially unacceptable, and recognizably so even without any kind of modern awareness of the issues.
|
# ? Jul 28, 2020 20:13 |
|
Like, if someone's name is listed as "Alex Smith" then maybe just, like, call them "Alex"? I'm legit confused in what context algorithmicly figuring out whether "Mr" or "Ms" Smith is more accurate beats just calling them "Alex"?
|
# ? Jul 28, 2020 20:14 |
|
Schadenboner posted:Like, if someone's name is listed as "Alex Smith" then maybe just, like, call them "Alex"? I'm legit confused in what context algorithmicly figuring out whether "Mr" or "Ms" Smith is more accurate beats just calling them "Alex"? presumably its just to pick what kind of ads to send them so the stakes are relatively low
|
# ? Jul 28, 2020 20:20 |
|
ultrafilter posted:
big johnny cash energy Schadenboner posted:Like, if someone's name is listed as "Alex Smith" then maybe just, like, call them "Alex"? I'm legit confused in what context algorithmicly figuring out whether "Mr" or "Ms" Smith is more accurate beats just calling them "Alex"? it’s so you can decide to serve them ads for pink razors or gray razors.
|
# ? Jul 28, 2020 20:21 |
|
hobbesmaster posted:presumably its just to pick what kind of ads to send them so the stakes are relatively low Like, I don't "do" marketing or ad-tech or whatever and I have uBlock Origin and PiHole so I don't interact with a lot of ads but It seems like using prior browsing habits might be a smarter way to drive engagement?
|
# ? Jul 28, 2020 20:22 |
|
Midjack posted:big johnny cash energy https://www.youtube.com/watch?v=5cYPEkpwi-0 https://www.youtube.com/watch?v=CWtcE52LKgk
|
# ? Jul 28, 2020 20:22 |
|
It's in front of an aircraft carrier. Because: of course?
|
# ? Jul 28, 2020 20:23 |
|
Sleng Teng posted:Dunno wtf "the team is working on separating the concepts of name/username/email from gender identity" youre the ones putting a gender identity on my email in the first place!!!
|
# ? Jul 28, 2020 20:23 |
|
Sagebrush posted:hm gently caress too late oh well
|
# ? Jul 28, 2020 20:30 |
|
Carthag Tuek posted:wtf "the team is working on separating the concepts of name/username/email from gender identity" i suspect they mean “separate” in the same way you do when you talk about distillation - extracting something useful from an aggregate.
|
# ? Jul 28, 2020 20:34 |
|
every site that sells ads has been doing this forever, like we had a proof of concept when I worked at evite 20 years ago. it wasn’t MaChInE LeArNiNg but it also didn’t think susan was a male name
|
# ? Jul 28, 2020 20:47 |
|
Sagebrush posted:
how would any of us even remember to loving breathe without the gentle guiding intellect of america's smartest boy
|
# ? Jul 28, 2020 21:01 |
|
carry on then posted:how would any of us even remember to loving breathe without the gentle guiding intellect of america's smartest boy always happy to help, carry on then
|
# ? Jul 28, 2020 21:03 |
|
|
# ? May 27, 2024 03:16 |
|
I suspect that was trained on a very unbalanced data set, as it's trivial to find simple, common words that spit out 99+% chance of being male, but "magenta" was the highest I got for female in my random samplings, and that was still just 95%. It has no conception of order being relevant to names, and will declare "Jennifer Douglas" and "Douglas Jennifer" are both 58% male / 42% female. Douglas wins that because "Douglas" alone is 99.1% likely to be male, but "Jennifer" is 98.1% likely to be female.
|
# ? Jul 29, 2020 00:02 |