[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/sci/ - Science & Math

View post   
File: 211 KB, 800x999, einstein_patentclerk.jpg [View same] [iqdb] [saucenao] [google]
7357829 No.7357829 [Reply] [Original]

Machine Learning is to the 2010s as Theoretical Physics was to 1905.

>> No.7357835

Network science is to the 2010s as theoretical physics was to 1905.

>> No.7357836

>>7357829
Maybe once machine learning comes a bit farther we'll start to see people stop treating intelligence as something human or even organic.

>> No.7357848

>>7357829
Theoretical Physics is to the 2010s as Theoretical Physics was to 1905.

>> No.7357852

>>7357829
And theoretical physics created the atomic bomb. Be careful what you wish for because you just might get it.

>> No.7357867

>>7357852
damn.. is he right?

>> No.7357932

>>7357836
The problem is that the qualities humans associate with intelligence (natural language processing, artistic, expression, awareness of mortality, etc) and the things that AI researchers work on (prediction of complex systems like weather or stocks, pattern recognition, etc) are totally different.

We aren't researching machine learning to make new humans.

>> No.7357953

>>7357829
>Machine Learning is to the 2010s as Theoretical Physics was to 1905.
Examples?

>> No.7358020

>>7357932
Yeah that is something that bothers me. I know people don't know any better but it gets annoying seeing pop sci articles and videos of "singularity" or some terminator type A.I scenario.

>>7357953
I think op means that it is a hot topic right now and there has been a push in researching ML and A.I right now. I don't think ML researchers are on the same magnitude of physicists but people like Micheal I Jordan, Daphne Koller, Zoubin Ghahrammi, Vapnik, and a personal favorite Botzheim have done some significant work in the past few years a kin to the physicists of that early 20th Century.

>> No.7358025

>>7358020
Not mentioning Andrew Ng.

For shame!

>> No.7358028

>>7358025
Oh shit. I was just going of the top of my head. I can't believe I forgot based Ng. How do I repent?

>> No.7358034

>>7357829
Maybe true but we still haven't seen any influence of it in industry. It's currently a bit of scientific curiosity.

>> No.7358050

>>7358020
Physics is easy

>> No.7358056

>>7358020
>I think op means that it is a hot topic right now and there has been a push in researching ML and A.I right now.
Yeah I get that, but what are the recent developments in machine learning that are comparable to special relativity in 1905.

>> No.7358061

Machine Learning is more decentralized than theoretical physics is though. There's not a few competing theories, there's a sprawling ecosystem of theories that build and borrow from each other. Hidden Markov Models are all the rage in finance right now, which is obviously being borrowed from Machine Learning.

>> No.7358063
File: 70 KB, 500x500, 1433716247384.jpg [View same] [iqdb] [saucenao] [google]
7358063

>>7357829
>comparing stuff with merit to probabilistic systems labelled with catchy misleading terms like 'learning' and 'intelligence'

>> No.7358067

>>7357932
It's not simply that people choose not to research those things. There was a great deal of interest in those areas decades ago and it was found that general intelligence was far more difficult to figure out than anyone thought. What you see now is the reaction to that.

>> No.7358070

>>7358063
This is a good point. A lot of "data science" and "machine learning" is just modified linear regression with a massive amount of data.

>> No.7358078

>>7358061
So is that good or bad news for Machine Learning?

>> No.7358082

>>7358078
>21st century
>decentralized
>is that good or bad
Do you even need to ask

>> No.7358086

I'm not sure what OP is implying.

>> No.7358090

>>7358086
Probably that current developments in machine learning are an historic moment.

>> No.7358096

>>7358082
But doesn't decentralization mean more healthy competition?

>> No.7358101

>>7357932

It's consciousness and I don't think we understand it, as far as I know. Until we figure it out, we can't have intelligence.

We can have really smart programs with the ability to learn, but being smart isn't being intelligent.

>> No.7358104

>>7357852
And it has prevented wars. So you're saying this is a good thing after all?

>> No.7358107

>>7358090
Like autonomous vehicles and facial/voice recognition? I see consumer applications to machine learning, but not much else. Granted, I do know next to nothing on the subject.

>> No.7358116

HOW DOES CAPTCHA KNOW WHAT FOOD IS!?!? THIS IS INSANITY.

>> No.7358119

>>7358096
>assuming implied implications incorrectly

>> No.7358166

Physics has been seriously hamster-wheeling for a long time now. I actually believe that humans are too limited by perspective or are not smart enough to unravel the true nature of reality and the next step in physics will require AI insight.

>> No.7358201

>>7358166
What precisely does machine learning have to do with A.I.?

>> No.7358213

>>7358063
>hiding behind undefined terms because you are too dumb to understand how the brain actually works.

>> No.7358214

>>7358201
Learning algorithms from Machine Learning will almost certainly be used in AI agent systems.

>> No.7358215

>>7358201
Machine learning refers to a set of algorithms, concepts, and techniques that are all a subset of artificial intelligence.

>> No.7358328
File: 27 KB, 527x409, 1413808723962.png [View same] [iqdb] [saucenao] [google]
7358328

>>7358213

>> No.7358338

>>7358201
it's in the name

>> No.7358344

>>7358201

Fair dinkum man, what the fuck is wrong with you?

>> No.7358364

Too mediocre math in the background. The most so-called "researchers" in this field never bother to go beyond playing around with standard MATLAB tools for fuzzy logic or neural netowrks. And these dudes are mostly the Indians and Kebabs (+ a bonus: the Africans)

>> No.7358409

>>7357932
>>7358020
Way to grossly misinterpret me.

>> No.7359231

>>7358328
Do you really think the brain is not a probabilistic system?

>> No.7359242

>>7358364
Not true man. I know why you have that impression, but serious machine learning is becoming very formalized and mathematical. That the scary part actually, they are much further ahead in this than people generally think.

Watch this lecture. They are using plenty of math
https://www.youtube.com/watch?v=kPrHqQzCkg0

>> No.7359253

>>7359242
Ugh no dude, I'm an ML grad and I can say that the math in ML, especially in the video you posted, is not as rigorous as you think.
>They are using plenty of math
>Maximum likelihood
>Deep learning
>Plenty
They are all basic linear algebra and applied statistic.
To be honest, the most rigorous part of ML is optimization and it's not even CS related.

>> No.7359259

>>7359253
Well what little math they are using seems to be working.

What are you getting at here? Do you think if they just make things more complex and add mathematical notation it will become legitimate?

>> No.7359277

>>7359259
I'm not saying that they are not working, It is legitimately using Math. But my point was that the Math being used in those models is not very ahead of other fields. That's all.
Of course we don't add complicated things into existing model just for the sake of it. Still, saying that they are much further ahead than people generally think is ridiculous.

>> No.7359283

>>7359277
And perhaps that leads you to the key insight behind machine learning.

You can use simple processes to model processes far more complex than themselves.

>> No.7359288

>>7359283
I agree with you on this one. However, current Deep net models are still pretty inefficient and it will need to be improved sooner or later. They also lack generalization and needs a lot of training samples (up to hundreds millions) to work.
Training a large scale Deep net is a fucking pain, even more so when you have to grid search the hyper params.

>> No.7359295

>>7359288
Right. I ultimately think these gigantic Back Propagation nets are a passing fad. People get excited because they are breaking world records on various benchmarks on a monthly or even weekly basis. But people need to start looking to other ways of training neural architectures.

>> No.7359297

>>7358338
>linear regression is the definition of intelligence

ok buddie.

>> No.7359304

>>7359297
It isn't? You weren't taught how to multiply? You weren't taught how to predict new Y given some example X and Y?
It's simple as fuck but it does the job. I bet that a lot of business fags using excel don't even know how to do linear regression anyway.

>> No.7359313

>>7358104
>the bomb has prevented war

Only direct full-blown nuclear wars ie things that wouldn't have been possible without the bomb anyway. Everything below that has happened despite nuclear dissuasion.

>> No.7359331

>>7358201
Detecting a pattern.

>> No.7359345

Is computer vision dead? What are the other cool AI subfields apart from ML? I am deciding my field for my grad course but I'm not sure.

>> No.7359346

>>7358107
Data processing generally. We live in the Eldorado of mass data, contrary to previous generations, our problem is not getting data point but making sense of them. Machine learning is one possibly powerful tool for this.

>> No.7359349

>>7358364
Don't worry bro, I'm ready to weigh him, armed with Holy Functional Analysis, Sacred Riemannian Geometry and Very Convenient Statistics.

>> No.7359405

ITT tryhard wannabes

>> No.7359408

No.

>> No.7359443

i'd compare it more to string theory

both extensively require new math to move forward

>> No.7359455

>>7359304
excel has a linear regression feature. I'm sure business fags using excel know how to click a button to get a line of best fit.

>> No.7361558

Machine learning has made next to no real progress in the last 45 years, other than progressing the science of academic grant vacuums to keep the business of holding knowledge hostage going.

>> No.7361565

>>7361558
>no real progress in the last 45 years
What about speech and image recognition?

>> No.7361569

>>7361565
Nope. Any marginal incremental benefits are on the hardware.

>> No.7361575
File: 33 KB, 625x626, 1435198572944.jpg [View same] [iqdb] [saucenao] [google]
7361575

>>7361569

>> No.7361584

>>7361575
Meme all you want, but unless you ask any academic in the field, you won't know how little we have progressed in the last decades.

When all you see is shit like "making an app for the iphone that uses an algorithm from the 60s with the added benefit of an OS capable of faking realtime speech recognition at higher input samples because newer processors are SUPERFAST!", you know something's wrong.

>> No.7361586

>>7361584
I am well aware, the current computational methods are just no suitable for AI.

real AI will come when a new medium that is non-computational in nature is developed.

but saying the field has made NO IMPROVEMENT
is a bit harsh

>> No.7361593

Computer science is a meme science. Discuss. [100]

>> No.7361640

>>7361586
>new medium that is non-computational
DARPA's Physical Intelligence program
rt.com/usa/systems-intelligence-robots-defense-781/

>> No.7362589

>>7361586
>I am well aware, the current computational methods are just no suitable for AI.

This is very debatable actually. Especially if you consider the possibility that the brains implementation of the algorithms which give you intelligence are not the most efficient possible.

But I do agree, using a brain inspired approach is the only sensible thing to do right now, and neural computing architectures will make that much easier.

>> No.7362882

>>7361584
>When all you see is shit like "making an app for the iphone that uses an algorithm from the 60s with the added benefit of an OS capable of faking realtime speech recognition at higher input samples because newer processors are SUPERFAST!", you know something's wrong.

Yeah idk if that's really a huge indictment.
Human language is like top-level of top level. It's probably the crowning achievement of our species, the epitome of evolution as we know it.