/sci/ - Science & Math » Thread #11391167

29KiB, 800x800, angry-emoji-1.jpg

View Same Google iqdb SauceNAO

Anonymous Sun 16 Feb 22:43:18 2020 No.11391167 View Reply Original Report

Quoted By: >>11391272 >>11391284 >>11391358 >>11392106 >>11392184 >>11393949 >>11394251 >>11394298 >>11394342

>machine learning / deep learning paper
>instead of explaining the model just put some random equations in the shittiest, most obscure possible notation.
Why do people do this? I have a feeling it's intentional to hide poor methodology.
Does this happen in other areas too?

Anonymous

Anonymous Sun 16 Feb 2020 23:31:18 No.11391272 Report

Quoted By: >>11391277 >>11391284

>>11391167
the whole idea of machine learning is that you don't care about what the fuck is going on, instead you feed your shit into the black box and poof, out comes your solution. so it is no surprise that they hand-wave away anything resembling human understanding

Anonymous

Anonymous Sun 16 Feb 2020 23:33:57 No.11391277 Report

Quoted By:

>>11391272
Will it be possible for deep learning machines to investigate each other and give us some easy human readable information?

Anonymous

Anonymous Sun 16 Feb 2020 23:36:22 No.11391284 Report

Quoted By: >>11391314 >>11391318 >>11391324

>>11391167
>>11391272
>t. never actually studied ml formally

Anonymous

Anonymous Sun 16 Feb 2020 23:49:46 No.11391314 Report

Quoted By: >>11391329 >>11392115

>>11391284
>t. studied ml formally
sorry for your poor life choices anon

Anonymous

View Same Google iqdb SauceNAO worried.png, 489KiB, 600x782

Anonymous Sun 16 Feb 2020 23:50:52 No.11391318 Report

Quoted By: >>11391329 >>11392115

>>11391284
>ml
>formally

Anonymous

View Same Google iqdb SauceNAO 0D4CC0B1-1C40-4AC2-A06E-B126C5E3 (...).png, 571KiB, 1120x1080

Anonymous Sun 16 Feb 2020 23:52:30 No.11391324 Report

Quoted By: >>11391329 >>11392115

>>11391284
>meme learning formally

Anonymous

Anonymous Sun 16 Feb 2020 23:54:36 No.11391329 Report

Quoted By: >>11391339 >>11391340 >>11391351 >>11391360 >>11394255

>>11391324
>>11391318
>>11391314

Can someone give me a rundown on machine learning. Why does sci hate ML and why is formal machine learning a meme too?

Anonymous

Anonymous Sun 16 Feb 2020 23:57:52 No.11391339 Report

Quoted By:

>>11391329
they never took a proper ml course. don't know abut how autistic it can get. they think it's just run this code and bam, results.

Anonymous

Anonymous Sun 16 Feb 2020 23:57:54 No.11391340 Report

Quoted By: >>11391350

>>11391329
because ML is the black box solution that replaces real understanding, and formal ML “studies” amount to basic linear algebra and machine learning with added nonlinear chaotic terms that are insoluble analytically. so basically it boils down to the formal study of intractable chaotic nonlinear systems to approximate curve fits, which nobody can do except numerically, which is what ML is

Anonymous

Anonymous Sun 16 Feb 2020 23:59:20 No.11391350 Report

Quoted By:

>>11391340
hidden layer doesn't mean what you think it means.

Anonymous

Anonymous Sun 16 Feb 2020 23:59:32 No.11391351 Report

Quoted By:

>>11391329
because ML is the black box solution that replaces real understanding, and formal ML “studies” amount to basic linear algebra and regression analysis with added nonlinear chaotic terms that are insoluble analytically. so basically it boils down to the formal study of intractable chaotic nonlinear systems to approximate curve fits, which nobody can do except numerically, which is what ML is

Anonymous

Anonymous Mon 17 Feb 2020 00:03:24 No.11391358 Report

Quoted By:

>>11391167
>machine learning / deep learning
>>>/g/

Anonymous

Anonymous Mon 17 Feb 2020 00:04:40 No.11391360 Report

Quoted By: >>11391369 >>11392164

>>11391329
Machine learning is not a meme. People who call it a meme are probably statisticians angry they're loosing their jobs or get paid less than the average machine learning position. Those black box models just work, and they're easy to implement and are making companies billions in revenue.
The thing is, even though it's not a meme, it does not have a formal theory to support it yet. The first person to formalize ML will be the richest person in history though. If you come up with a serious theory of learning/generalization and whatnot that allows you to predict how well neural networks in general will do for a given problem and other such things, people will literally give you rivers of gold.
What people here are calling "formal ML" is either just differential geometry memes that can't be used to predict shit beyond optimal cases, or they're just bluffing and hoping no one calls them on it.

Anonymous

Anonymous Mon 17 Feb 2020 00:08:26 No.11391369 Report

Quoted By: >>11391376

>>11391360
ML has a formal theory to it. it's called "neural networks" and guys like Parisi and L. Cooper worked on it for decades. the issue is that the formal theory leads to insoluble linear equations that can only really be solved if you invoke some "tricks" like they use in spin glass physics. none of this applies to the advertisement-algorithm crap that the tech industry makes billions on, however.

maybe somebody might crack the theory but i doubt it would make anyone besides the tech industry any billions. heck, if i discovered it today, the intellectual property would either belong to the public domain or my university, not me

Anonymous

Anonymous Mon 17 Feb 2020 00:09:51 No.11391376 Report

Quoted By:

>>11391369
*insoluble nonlinear equations
not "insoluble linear equations"

Anonymous

View Same Google iqdb SauceNAO 1579322327178.jpg, 15KiB, 543x640

Anonymous Mon 17 Feb 2020 06:43:19 No.11392106 Report

Quoted By:

>>11391167
>deduction paper
>instead of explaining their approach they just give a new logic with the most obscure notation possible
>mfw

Anonymous

Anonymous Mon 17 Feb 2020 06:51:15 No.11392115 Report

Quoted By:

>>11391324
>>11391318
>>11391314
the absolute state of this board

Anonymous

Anonymous Mon 17 Feb 2020 07:34:38 No.11392164 Report

Quoted By:

>>11391360
This. But ML does have a formal foundation, statistical learning theory. What you meant is DL.

Anonymous

Anonymous Mon 17 Feb 2020 07:50:09 No.11392184 Report

Quoted By: >>11394293

>>11391167
if you want actual rigor friend, then read computational learning papers, since that's where the theory activity tends to concentrate. ML papers are usually just focused on "apply X technique to Y problem in a novel approach where we change Z."

Anonymous

Anonymous Mon 17 Feb 2020 20:41:29 No.11393949 Report

Quoted By:

>>11391167
Machine Learning is interesting

Anonymous

Anonymous Mon 17 Feb 2020 22:19:40 No.11394251 Report

Quoted By:

>>11391167
Link to paper?
>I have a feeling it's intentional to hide poor methodology.
Sometimes yes, sometimes it's because of how nureviewers behave (if you don't shoehorn math you don't need/want in your paper, "not enough math rejected!", if you put in obscure math that's barely relevant (and usually completely wrong for bonus points), "there's maths that means it's legit [accepted]").
>Does this happen in other areas too?
Literally, LITERALLY everywhere.

Anonymous

Anonymous Mon 17 Feb 2020 22:21:05 No.11394255 Report

Quoted By:

>>11391329
Because one senile old fart says he doesn't like it and since said old fart was a mathematician, that means he's totally not yet another advent of black science man and they should parrot the dumb shit he's saying like it's the very word of God.

Anonymous

Anonymous Mon 17 Feb 2020 22:35:45 No.11394293 Report

Quoted By:

>>11392184
In theory you're correct (haha) you're correct, but in practice nobody uses these terms like that.
For example:

https://arxiv.org/pdf/1706.03471.pdf
https://arxiv.org/pdf/1307.2118.pdf
http://www.jmlr.org/papers/volume2/bousquet02a/bousquet02a.pdf
https://arxiv.org/pdf/1509.01240.pdf

These papers are all identified as pure machine learning paper, despite being at the edge of the theoretical advances of machine learning for their time.

Anonymous

Anonymous Mon 17 Feb 2020 22:37:07 No.11394298 Report

Quoted By:

>>11391167
>deep learning *******paper*******
you brought this to yourself

Anonymous

Anonymous Mon 17 Feb 2020 22:51:47 No.11394342 Report

Quoted By:

>>11391167
Yes, they do this to hide poor methodology.
Yes, it does happen in other areas too.
No, most people in this thread don't know shit about ML/DL. Even though learned parameters are not easily interpretable, good ML/DL researchers should be able to discuss dataset, activation functions, network architecture, results, etc. as well as their implications and how they relate to each other.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches