Algorithms define our lives
November 6, 2018 6:11 AM   Subscribe

 
Where we have seen breakthroughs is in the results of applying huge quantities of data to flexible models to do very particular tasks in very particular environments. The systems we get from this are really good at that, but really fragile
...
artificial intelligence is the technology of the future, and always will be.
(This is good)
posted by Mayor West at 6:49 AM on November 6 [4 favorites]


It's an important topic and all, but still I can't help it. Every time people start talking about "algorithms" and their racial bias, discrimination, their embedded values and power for social control, the danger that they might influence elections, I think back to Computer Science 103: Algorithms and imagine that they're talking about quicksort or something.
posted by sfenders at 6:50 AM on November 6 [14 favorites]


Bumblesort
posted by pracowity at 6:59 AM on November 6 [1 favorite]


Thanks, I really liked the first article after the fold. It's a clear and concise survey of some of the central issues in deep learning and AI.

Though it seems odd to me that the author says that 90% of machine learning is nonparamatric regression, when I would consider deep learning models to be parametric, and also constitute a large part of machine learning? Although I did see a talk recently where the speaker argued that deep learning models are simply memorizing datasets and interpolating, which is essentially what nonparametric models do.
posted by Alex404 at 7:01 AM on November 6 [1 favorite]


Yeah, the use of the word “algorithms” for this in the public imagination is weird and unfortunate. “Models” or “statistical models” would have been better.

I guess “algorithms” must have gotten better clickthrough when they were A/B testing headlines for the first bunch of articles about this a couple years ago.
posted by vogon_poet at 7:22 AM on November 6 [7 favorites]


I’m not sure what “nonparametric” means in this context. Deep nets obviously have a bunch of parameters. I’ve seen “nonparametric” used for models that effectively have an adaptable number of parameters, so it could be something to do with that. A lot of non-deep machine learning does fall into that category.
posted by vogon_poet at 7:28 AM on November 6


> Computer Science 103: Algorithms and imagine that they're talking about quicksort or something.

Quicksort is fine. Naïve Bayes is the one we really gotta worry about.

though let me tell you I’ve never trusted that Floyd-Warshall.
posted by Reclusive Novelist Thomas Pynchon at 7:44 AM on November 6 [4 favorites]


Are algorithms discovered or invented? Karl Marx went to great lengths to describe the consequences of capitalism, an "AI" composed of nothing more than gold and men's greed. Darwin likewise described an "AI" that was ultimately shown to reside in minute chemicals.
posted by SPrintF at 7:59 AM on November 6 [1 favorite]


Cathy O'Neil's short video - "an algorithm is data, and a definition of success" - is perfectly illustrated by Amazon's hiring-algorithm fuckup. The algorithm looked at historical hires to figure out the best current applicants, and the algorithm wound up learning that men were the "best" candidates. Because in the past, mostly men had been hired.

So "who's involved in tech" winds up meaning "who's defining success" and as tech products and services increasingly define our lives and our economy, these algorithms' definition of success becomes pretty damn important.
posted by entropone at 8:03 AM on November 6 [4 favorites]


Mathbabe!

(don't jump on me, that's Cathy's blog, her title. Great blog, thought it'd gone quiet but I see there are quite amusing recent updates) (and not just math and data science vs ethics, but a whole ranges of, cough, omgveryinteresting topics)
posted by sammyo at 8:14 AM on November 6 [1 favorite]


Yeah, the use of the word “algorithms” for this in the public imagination is weird and unfortunate. “Models” or “statistical models” would have been better.

Perhaps there is a better vocabulary, but it is too late now.
posted by thelonius at 8:32 AM on November 6 [1 favorite]


Quicksort is fine. Naïve Bayes is the one we really gotta worry about.

Topological Sort or go home to flatland
posted by sammyo at 9:02 AM on November 6 [1 favorite]


Flatland be buggered. Ima sit here in the shade of these k-d trees until it becomes obvious where it's going to be easiest to get to next.
posted by flabdablet at 9:08 AM on November 6 [2 favorites]


> Topological Sort or go home to flatland

really the problem lately is that we’ve been having trouble finding our strongly connected components.
posted by Reclusive Novelist Thomas Pynchon at 9:10 AM on November 6


Are nested if statements going to take our jobs?

Why isn't the Government doing more about .jpgs?

Is your child using strings?

What if China gets a viable installation of Python?
posted by Damienmce at 9:16 AM on November 6 [8 favorites]


okay everyone the challenge for the rest of this thread is to try to make every subsequent comment both a serious statement about contemporary politics and also a very silly math joke. Why should we do this? Because that is exactly what Silicon Valley has done to our field of political discourse. The most influential political statement is now the one that’s most easily decomposable into clever math.
posted by Reclusive Novelist Thomas Pynchon at 9:18 AM on November 6


Perhaps there is a better vocabulary, but it is too late now.

Not really? In the working world, we talk about machine learning algorithms which produce models. I mainly see the unmodified word "algorithms " in classes that teach quicksort and articles that clutch pearls....

I'm a bit worried that Cathy O'Neill seems to be embracing less precise language. Algorithms are not data and a goal; an algorithm is a method or process for doing something. And encouraging blind opposition to 'algorithms' is pretty misguided, imho; the status quo, for those who haven't been paying attention, also sucks. It's important to work on fairness and oversight.

Consider Equifax. They suck. Switching to new systems with fairness and oversight built in could give us something else that sucks, but is tweakable. But the status quo are credit bureaus which are unaccountable, unchangeable, unfair and often just wrong.

There's a statement I like at the end of the IMF article: that for the near future, the main economic impact of ML/AI is going to be businesses making their systems more machine compatible. This would be a fantastic outcome: the same work will make it much easier to audit and regulate these business systems.
posted by kaibutsu at 9:23 AM on November 6 [1 favorite]


>> "Perhaps there is a better vocabulary"

How 'bout "Systemic Discrimination"?
posted by spudsilo at 9:25 AM on November 6 [4 favorites]


I mainly see the unmodified word "algorithms " in classes that teach quicksort and articles that clutch pearls....

It's the latter usage which will be, I predict, impossible to change.
posted by thelonius at 9:26 AM on November 6


how many mathematicians does it take to screw in a lightbulb?
a: they cite three californians, reducing it to a previous joke.

how many machine learning algorithms does it take to screw in a lightbulb?
a: possibly one, but it will require give me $20MM in VC funding and will eventually pivot into a gig economy lightbulb screwing app.
posted by kaibutsu at 9:27 AM on November 6 [4 favorites]


How 'bout "Systemic Discrimination"?

How about "Risks to the Public in Computers and Related Systems"? The "hiring-algorithm fuckup" related above is exactly the sort of news they've been covering since the 1980's.
posted by sfenders at 9:35 AM on November 6


systems more machine compatible. This would be a fantastic outcome: the same work will make it much easier to audit and regulate these business systems.

To keep with the math theme, this is not just a double edged sword, consider n edges: n=3 is a stiletto, a wound that does not close. n=20 ouch. n=∞ ok silly, but data will be used in ways that are just that many dimensional, many purposes, many helpful and incredibly useful -- we don't want to go backwards but there are many bad possibilities.

Really need legislation at least at the bill of rights level, data capture is not going to stop and the tools to analyse will grow in power and NO ONE, no researcher, no expert, no where no how, at any level ,has any actual functional idea how to wrangle this problem.
posted by sammyo at 10:07 AM on November 6 [1 favorite]


> There's a statement I like at the end of the IMF article: that for the near future, the main economic impact of ML/AI is going to be businesses making their systems more machine compatible. This would be a fantastic outcome: the same work will make it much easier to audit and regulate these business systems.

Well but also, the processes by which we live our day-to-day lives are thereby also becoming more auditable, regulatable, and machine-compatible, which can allow us to see new efficiencies but that also can also result in us presenting new attack surfaces for new and subtle types of oppression. See James C. Scott for arguments that run roughly along these lines...
posted by Reclusive Novelist Thomas Pynchon at 10:09 AM on November 6 [1 favorite]


My wife's insulin pump has an auto mode that uses an algorithm to make minor adjustments to your insulin delivery. The FDA (smartly) didn't sign off on the pump acting like an artificial pancreas although clearly, that is the direction all this is going. My wife tried it for one day and noped right back to manual mode. It was treating her as low at a blood glucose level over 100 and cutting off her insulin supply. If you don't control your type I diabetes a machine that wants to keep you at 180 is a definite improvement to your health, but that is a huge step back for somebody with good control.

We asked about the algorithm and of course, Medtronic won't release any information about how it works.

Personally, I think any algorithm controlling health like that should be open source or public domain. We all have the right to know exactly how a medical device is making decisions about our health.
posted by COD at 11:00 AM on November 6 [9 favorites]


At some point, it’s Sapir-Whorf, all the way down.
posted by jenkinsEar at 11:20 AM on November 6 [1 favorite]


> Personally, I think any algorithm controlling health like that should be open source or public domain. We all have the right to know exactly how a medical device is making decisions about our health.

In the final analysis it's impossible to fully trust any piece of software, even if you have the source code.

but on a more prosaic level, yes. There's been serious medical disasters that would have been prevented had more eyes been on the source code used in medical devices.

(my "favorite" of the lethal Therac-25 bugs is that it incremented a counter whenever it encountered an error during setup, and refused to proceed with irradiation if that counter was non-zero. except the counter was eight bits wide, meaning that everything works perfectly... until the run where you encounter exactly 256 errors during setup.)
posted by Reclusive Novelist Thomas Pynchon at 11:31 AM on November 6 [1 favorite]


Part of my team is working on a big data project (I would bet some money that most regular readers of mefi have some of their online activity captured in the data set) with a complicated client.

The non technical biz people call everything 'the algorithm', as in 'someone broke the algorithm, YouTube is slow on my computer'. The more technical people talk about models and networks and data, and know exactly were the algorithm heavy work goes.

The problem is that it is the biz people who dictate the assumptions to be baked into the models.

The goal set by the biz is 'revenue' and nothing else.

On their free time my team sets goals like 'interesting' and 'insigthful', and they find really cool stuff in the data that has absolutely no value for the client, but which would very very interesting for an anthropologist or someone into policy making.

In my ideal world all data sets are free and the people are educated in data science. In the meantime data sets belong to corporations and black hat hackers, and math is hard.
posted by Dr. Curare at 11:52 AM on November 6 [6 favorites]


I think Montoya's Remark is appropriate here.
posted by erniepan at 6:32 PM on November 6


classes that teach quicksort and articles that clutch pearls....

It's the latter usage which will be, I predict, impossible to change.


Oh it will change, all right. You just won't like the results.

See also: "life hack".
posted by flabdablet at 10:49 PM on November 6


« Older Worst. Ranked list. EVAR.   |   Happy Birthday Eugene! Newer »


You are not currently logged in. Log in or create a new account to post comments.