1838: "Machine Learning"

This forum is for the individual discussion thread that goes with each new comic.

Moderators: Moderators General, Prelates, Magistrates

User avatar
zanglebert
Posts: 12
Joined: Thu Jun 26, 2014 6:31 am UTC

1838: "Machine Learning"

Postby zanglebert » Wed May 17, 2017 8:47 am UTC

Image

Title text = The pile gets soaked with data and starts to get mushy over time, so it's technically recurrent.

Clever presentation, but I'm a bit unimpressed by the same old message. Some grain of truth in it, I guess, but ML doesn't exactly work like that. Should have asked Hinton, Bengio, Goldberg or any of the other guys and girls who know what they're doing and why.

User avatar
Eternal Density
Posts: 5479
Joined: Thu Oct 02, 2008 12:37 am UTC
Contact:

Re: 1838: "Machine Learning"

Postby Eternal Density » Wed May 17, 2017 11:05 am UTC

That's sorta a high level outline of some of the things involved, but it's not very informative.
I like this article http://karpathy.github.io/2015/05/21/rnn-effectiveness/
My favourite application of an RNN is https://twitter.com/RoboRosewater
And of course the classic card: (spoilered for large image)
Spoiler:
Image

For anyone who cannot see the image, just look up Slidshocking Krow. (or Tromple and Mointainspalk)
Play the game of Time! castle.chirpingmustard.com Hotdog Vending Supplier But what is this?
In the Marvel vs. DC film-making war, we're all winners.

User avatar
rhhardin
Posts: 51
Joined: Fri Apr 09, 2010 2:11 pm UTC

Re: 1838: "Machine Learning"

Postby rhhardin » Wed May 17, 2017 11:43 am UTC

We had the Chaostron in the 50s, constructed with random Western Electric wiring diagrams and emulated (until funds ran out) on an IBM 709.

V.A.Vyssotsky manually completed the partial last simulation.

It printed a line of punctuation and ejected the printer paper twice, an early sign of intelligence.

User avatar
Soupspoon
You have done something you shouldn't. Or are about to.
Posts: 2549
Joined: Thu Jan 28, 2016 7:00 pm UTC
Location: 53-1

Re: 1838: "Machine Learning"

Postby Soupspoon » Wed May 17, 2017 12:29 pm UTC

++?????++ Out of Cheese Error. Redo From Start.

User avatar
zanglebert
Posts: 12
Joined: Thu Jun 26, 2014 6:31 am UTC

Re: 1838: "Machine Learning"

Postby zanglebert » Wed May 17, 2017 1:13 pm UTC

Eternal Density wrote:That's sorta a high level outline of some of the things involved, but it's not very informative.
I like this article http://karpathy.github.io/2015/05/21/rnn-effectiveness/
My favourite application of an RNN is https://twitter.com/RoboRosewater
And of course the classic card: (spoilered for large image)
Spoiler:
Image

For anyone who cannot see the image, just look up Slidshocking Krow. (or Tromple and Mointainspalk)


These cards are amazing. I mean, I knew about the fake Shakespeare/fake research articles written in fake Latex, but those are new to me. Thanks for pointing them out.

And that's what we get by a simple RNN (with LSTM gates). A character-level RNN, I'd like to add... What I mean is, there's still so much room for activities architectural improvements, I can't wait to see what the next years will bring.

User avatar
ManaUser
Posts: 284
Joined: Mon Jun 09, 2008 9:28 pm UTC

Re: 1838: "Machine Learning"

Postby ManaUser » Wed May 17, 2017 4:01 pm UTC

Machine Learning may be good at alot of things, but clearly game design is not one of them. That card is freakin' OP.

a9s
Posts: 12
Joined: Wed Jan 21, 2015 7:06 pm UTC

Re: 1838: "Machine Learning"

Postby a9s » Wed May 17, 2017 5:42 pm UTC

Eternal Density wrote:That's sorta a high level outline of some of the things involved, but it's not very informative.
I like this article http://karpathy.github.io/2015/05/21/rnn-effectiveness/
My favourite application of an RNN is https://twitter.com/RoboRosewater
And of course the classic card: (spoilered for large image)
Spoiler:
Image

For anyone who cannot see the image, just look up Slidshocking Krow. (or Tromple and Mointainspalk)

Hands down the best one, Horse:
Spoiler:
Image

User avatar
CorruptUser
Posts: 8881
Joined: Fri Nov 06, 2009 10:12 pm UTC

Re: 1838: "Machine Learning"

Postby CorruptUser » Wed May 17, 2017 10:26 pm UTC

Blegh, I dislike "machine learning". I'm sure it's actually something useful to computer scientists, but from my experience it's mostly the new "quantum" in that no one outside the field really understands it but thinks it will just magically transform the world or let you do whatever scifi crap you want.

Stargazer71
Posts: 82
Joined: Mon Aug 23, 2010 3:00 pm UTC

Re: 1838: "Machine Learning"

Postby Stargazer71 » Thu May 18, 2017 2:39 am UTC

Do I dare to say that this is basically climate models in a nutshell? I think I will.

Not only because I love watching doomsday evangelists' heads explode, but also because it makes this joke eerily similar to Sunday's Dilbert (which also made doomsday evangelists' heads explode): http://dilbert.com/strip/2017-05-14

User avatar
Soupspoon
You have done something you shouldn't. Or are about to.
Posts: 2549
Joined: Thu Jan 28, 2016 7:00 pm UTC
Location: 53-1

Re: 1838: "Machine Learning"

Postby Soupspoon » Thu May 18, 2017 3:12 am UTC

Except that I've lost a lot of respect for Scott Adams, recently, and I'm trying not to let it influence my critical viewpoint, but whilst Randall is quite legitimately questioning the general process as a low-effort solution to anything and everything, Scott appears to be questioning the results obtained by a far more rigorous and precise methodology in the name of rubbishing those who 'believe' the results.

Now, because of my preconceptions, I might just be missing a deeper joke in the Dilbert that's actually funny, and it's not a strawman attack on climate non-deniers.... I'll have to come back to it later and consider it further.

moops
Posts: 10
Joined: Thu Aug 11, 2011 10:03 pm UTC

Re: 1838: "Machine Learning"

Postby moops » Thu May 18, 2017 4:39 am UTC

...no, I don't think there is a deeper joke here. It sort of goes "economics is bad at predicting, so climate predictions are bunk". Which I think Dilbert fans consider a kind of trenchant and witty insight, instead of a kind of rhetorical poop flinging.

moops
Posts: 10
Joined: Thu Aug 11, 2011 10:03 pm UTC

Re: 1838: "Machine Learning"

Postby moops » Thu May 18, 2017 4:41 am UTC

As for Machine Learning, I think they need to take a serious moment to think about p-hacking. The ML people I know have no clue what I'm talking about, which is worrisome.

User avatar
zanglebert
Posts: 12
Joined: Thu Jun 26, 2014 6:31 am UTC

Re: 1838: "Machine Learning"

Postby zanglebert » Thu May 18, 2017 8:06 am UTC

moops wrote:As for Machine Learning, I think they need to take a serious moment to think about p-hacking. The ML people I know have no clue what I'm talking about, which is worrisome.


Now that is a valid criticism of the field... "Our results improve the previous state-of-the-art on the task by 0.025%, which we consider to be a significant improvement. (But we didn't actually check that, duh!)"

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 25823
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1838: "Machine Learning"

Postby gmalivuk » Thu May 18, 2017 12:49 pm UTC

Stargazer71 wrote:Do I dare to say that this is basically climate models in a nutshell? I think I will.

"Do I dare out myself as a complete and total moron? I think I will."

No one's head will explode just from seeing another idiot science denier who thinks Dilbert provides insightful commentary on the topic.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

User avatar
Copper Bezel
Posts: 2416
Joined: Wed Oct 12, 2011 6:35 am UTC
Location: Web exclusive!

Re: 1838: "Machine Learning"

Postby Copper Bezel » Thu May 18, 2017 1:24 pm UTC

Soupspoon wrote:Except that I've lost a lot of respect for Scott Adams, recently, and I'm trying not to let it influence my critical viewpoint, but whilst Randall is quite legitimately questioning the general process as a low-effort solution to anything and everything, Scott appears to be questioning the results obtained by a far more rigorous and precise methodology in the name of rubbishing those who 'believe' the results.


I'm not even really reading 1838 as a serious criticism. If you describe machine learning in baby words, it kinda sounds like the image in the comic. Which is inherently funny to me.
So much depends upon a red wheel barrow (>= XXII) but it is not going to be installed.

she / her / her

User avatar
gmalivuk
GNU Terry Pratchett
Posts: 25823
Joined: Wed Feb 28, 2007 6:02 pm UTC
Location: Here and There
Contact:

Re: 1838: "Machine Learning"

Postby gmalivuk » Thu May 18, 2017 1:43 pm UTC

Yeah, this is a character who completely misses the mark on machine learning and has come up with this ridiculous thing as a result. People responding as though Randall is criticizing machine learning as a whole are in turn completely missing the mark on the comic.
Unless stated otherwise, I do not care whether a statement, by itself, constitutes a persuasive political argument. I care whether it's true.
---
If this post has math that doesn't work for you, use TeX the World for Firefox or Chrome

(he/him/his)

User avatar
pogrmman
Posts: 338
Joined: Wed Jun 29, 2016 10:53 pm UTC
Location: Probably outside

Re: 1838: "Machine Learning"

Postby pogrmman » Thu May 18, 2017 2:12 pm UTC

I really enjoyed this comic!

I don't think he's criticizing ML in general -- it's as if this character was trying to teach himself, saw all the linear algebra, didn't understand what was going on, threw a bunch of random linear algebra together, and decided "meh, good enough".

ML is a fascinating subject. I mean, I understand how neural nets and stuff work, but it still seems like black magic to me. The article on RNNs in one of the earlier replies (this one:
Eternal Density wrote:I like this article http://karpathy.github.io/2015/05/21/rnn-effectiveness/
) seems to demonstrate the sort of "black magic" aspect well. I mean, I understand what the net is doing, but it blows my mind that a character level RNN can produce the stuff on that page!

It does seem like a potential issue is the massive datasets -- so if you are trying to draw some sort of conclusion from them, you get results that appear to be significant when they aren't. This isn't much of a problem for task based ML, but we should still keep it in mind.

User avatar
Soupspoon
You have done something you shouldn't. Or are about to.
Posts: 2549
Joined: Thu Jan 28, 2016 7:00 pm UTC
Location: 53-1

Re: 1838: "Machine Learning"

Postby Soupspoon » Thu May 18, 2017 2:33 pm UTC

(Noting that I was talking about the low-effort "jumble things through a total randomness of anything" as what was being questioned by Randall... I'm currently designing a two-phase process to procedurally generate patterns to a given input in the first, which then get pushed through a recognition system in the second. The former is just genetically picking up archetype of building blocks, the idea being to look for a "low energy" and low-ambiguity way to represent the input. I don't worry much about the components I make available, because I plan to change it later. But to assess the relative 'fitness' I'm aware that the second phase needs to be designed-but-not-designed sufficient to be useful enough to give the kind of feedback I would want to push back if I had the time to do it myself (or crowd-source, in the manner oftne classic Darwinian Poetry), by seeing how easy it is to (re)train the latter to interpret each example of the former. Random functions without an eye for structural usefulness would make it an awful Learning Machine, but also have to strike away from preconceived notions for the processing. And this is just for fun!...Hence why I want to background it all, not spend my (or others') time staring at outputs...)

Mikeski
Posts: 876
Joined: Sun Jan 13, 2008 7:24 am UTC
Location: Minnesota, USA

Re: 1838: "Machine Learning"

Postby Mikeski » Thu May 18, 2017 11:46 pm UTC

gmalivuk wrote:Yeah, this is a character who completely misses the mark on machine learning and has come up with this ridiculous thing as a result. People responding as though Randall is criticizing machine learning as a whole are in turn completely missing the mark on the comic.

The guy with the shovel should probably have a white beret on.

Although then he wouldn't be getting any wrong answers...

niauropsaka
Posts: 72
Joined: Mon Apr 22, 2013 8:50 pm UTC

Re: 1838: "Machine Learning"

Postby niauropsaka » Fri May 19, 2017 5:28 am UTC

moops wrote:...no, I don't think there is a deeper joke here. It sort of goes "economics is bad at predicting, so climate predictions are bunk". Which I think Dilbert fans consider a kind of trenchant and witty insight, instead of a kind of rhetorical poop flinging.


Except that economic models aren't so much "bad at predicting" as "not actually followed by politicians (almost none of whom are economists) nor capable of restraining determined wrongdoers in the finance sector." Seriously, Simon Wren-Lewis and Paul Krugman have basically spent the last ten years yelling into the wind about this. The academic economists had had the models, they knew what they were doing, and then the politicians and bankers would ignore them and go with their guts instead.

It's a whole thing.

User avatar
Eternal Density
Posts: 5479
Joined: Thu Oct 02, 2008 12:37 am UTC
Contact:

Re: 1838: "Machine Learning"

Postby Eternal Density » Sun May 21, 2017 3:15 am UTC

The Magic card RNN development and discussion went on for a long time with lots of interesting diversions, over here http://www.mtgsalvation.com/forums/comm ... ent-neural

There's also a word-level rnn that generates clickbait:
https://larseidnes.com/2015/10/13/auto- ... -networks/

A while back I trained a char rnn on various text from this very forum, with amusing results.
Play the game of Time! castle.chirpingmustard.com Hotdog Vending Supplier But what is this?
In the Marvel vs. DC film-making war, we're all winners.

rmsgrey
Posts: 3116
Joined: Wed Nov 16, 2011 6:35 pm UTC

Re: 1838: "Machine Learning"

Postby rmsgrey » Sun May 21, 2017 8:33 pm UTC

niauropsaka wrote:
moops wrote:...no, I don't think there is a deeper joke here. It sort of goes "economics is bad at predicting, so climate predictions are bunk". Which I think Dilbert fans consider a kind of trenchant and witty insight, instead of a kind of rhetorical poop flinging.


Except that economic models aren't so much "bad at predicting" as "not actually followed by politicians (almost none of whom are economists) nor capable of restraining determined wrongdoers in the finance sector." Seriously, Simon Wren-Lewis and Paul Krugman have basically spent the last ten years yelling into the wind about this. The academic economists had had the models, they knew what they were doing, and then the politicians and bankers would ignore them and go with their guts instead.

It's a whole thing.


There was also an issue with a number of mainstream universities purely teaching orthodox economic models that predicted that 2008 couldn't happen - about 5 years ago, there were headlines about economics students going on strike because their teachers wouldn't cover any of the theories that could (and did) predict it.

User avatar
Steve the Pocket
Posts: 646
Joined: Mon Apr 23, 2007 4:02 am UTC
Location: Going downtuuu in a Luleelurah!

Re: 1838: "Machine Learning"

Postby Steve the Pocket » Mon May 22, 2017 5:50 pm UTC

Eternal Density wrote:There's also a word-level rnn that generates clickbait:
https://larseidnes.com/2015/10/13/auto- ... -networks/

I'm fairly certain the infamous "Does Bruno Mars Is Gay?" article was produced by a computer of some kind. It sure as hell fails the Turing Test at least.

On the frontiers of machine learning, we've got a couple great recent items: one that tries to generate D&D spells, and one that tries to generate paint colors. Methinks that second one would be especially of interest to people who remember Randall's color survey.
cephalopod9 wrote:Only on Xkcd can you start a topic involving Hitler and people spend the better part of half a dozen pages arguing about the quality of Operating Systems.

Baige.


Return to “Individual XKCD Comic Threads”

Who is online

Users browsing this forum: No registered users and 49 guests