/sci/ - Science & Math

[Anonymous No.7334216]

Where are all the machine learning people at?

[Anonymous No.7334229]

they're machines, not people, dumbass

[Anonymous No.7334233]

They're all busy making ridiculous salaries

[Anonymous No.7334302]

>>7334233
This.

[Anonymous No.7334373]

I love neural networks, but I haven't got the time to go into recurrents yet (mostly because the ML in coursera didn't go into it).

But I have a hard time finding problems that can be solved by it in my field. Aside from computer vision, what other applications does it have?

[Anonymous No.7334388]

bump for interest. I plan on going to grad school for ML. Anyone here been in a CS phd program?

[Anonymous No.7334393]

Good question.

Another (related) questions: is machine learning an option for a math major ? I have been told I should learn a language that's close to the machine, and stick with it, but besides that ?

Also what are the mathematical basis of machine learning ? I heard statistics were essential ?

[Anonymous No.7334417]

>>7334216
It's a very obscure field that most people don't even know exists. You aren't going to find very many of them on /sci/.

>>7334393
Learn calculus 2 and 3. Learn matrix algebra. That's pretty much it.

[Anonymous No.7334418]

>>7334388
I'm in a masters program focusing on machine learning, if that can help.

>>7334393
Math majors can absolutely get into machine learning. Programming languages really don't matter, but since you ask I'll list some that are used often: R, Matlab, Python (SciPy), LISP, and others. Statistics and combinatorics are a must; some areas will require different studies in each, including other subjects entirely.

>>7334373
Neutral networks essentially fit an n-dimensional curve. You can use them for solving any related problem depending on how the network is set up (objective function, etc.). These can include: classification, regression, belief state construction, object detection, value function approximation, optical character recognition, clustering, and others. Note that while neutral networks work for all of these, they may not be the best way to solve any particular problem.

[Anonymous No.7334423]

>>7334417
This is wrong.

[Anonymous No.7334426]

>>7334216
>machine learning
Is that what retards call computer science? Machine learnin'

[Anonymous No.7334427]

>>7334418
What is the program like ( difficulty and material)? What are you working on? What was your undergrad and where ? I just finished my first year of CS and minor in Stats so I really don't know much.I'm at a shitty state school so I am trying to make up by working in a research group at uni working on data mining. What are some absolute concepts I should know before going to grad school? thanks

[Anonymous No.7334436]

>>7334417
>It's a very obscure field that most people don't even know exists. You aren't going to find very many of them on /sci/.
attend any computer science class at a major university and you'll find people who are "into" machine learning. whether they know shit about it is another matter, but probably at least one or two will.

[Anonymous No.7334450]

>>7334427
Since I'm only a masters student, I still take courses along with my research credits. This means that even though I have 6 credit hours of research, I still spend a lot of time working outside of the lab. The main difference I can note between undergraduate and graduate classes is that conceptually the material requires more work and individual assignments are merely reinforcement of knowledge, and not the grunt work found in undergraduate classes.
My work is in model learning within the context of decision processes. That is, learning dynamics and relations within a goal directed process (in this case, stochastic games).
I got my bachelors at a large state university on the west coast in Computer Science. I now go to a different, but comparable university, with a larger research department in my area of interest.
You should note that the biggest concern when choosing a graduate school is whether or not they produce quality research in what field you want to work in.
You should go to your library and pick up a copy of Artificial Intelligence: A Modern Approach and read it front to back. Once done, go to your faculty website and find a professor with a research lab in what you are interested in, schedule an appointment, discuss with them your interests, and ask if they have an open research assistant position. This puts you around a lot of graduate students and may end up getting you published as an undergraduate (it's what I did).

[Anonymous No.7334472]

>>7334418
Thanks for your answer.


> Programming languages really don't matter, but since you ask I'll list some that are used often: R, Matlab, Python (SciPy), LISP, and others.

Okay, that's reassuring. I've some grasp of R and Matlab.

>Statistics and combinatorics are a must

I happen to love those.

> some areas will require different studies in each, including other subjects entirely.

I'm all for being ecclectic anyway, learning new maths doesn't scare me.

>tfw you realize you could have a decent job after all

Thanks /sci. You've been pretty helpful this time.

[Anonymous No.7334479]

>>7334472
No problem. I highly suggest giving the MIT OCW lectures on artificial intelligence a glance over.

[Anonymous No.7334484]

>>7334450
Thanks for the info.I have looked at some schools so I have a good idea where I would like to go to.
>Artificial Intelligence: A Modern Approach

will do thanks. There is one professor at my school that works in my interests so I'll look into that.Is publishing absolutely needed? I will make it my goal but I imagine it is pretty tough for an undergrad to get published. Are there really that many published undergraduates applying to grad school? Another option is getting an internship in data science to buff my app. Would that be able to make up for not being published?

[Anonymous No.7334499]

>>7334484
Being published isn't required, but it definitely helps. The internship might help a bit, but demonstrating that you can produce good research is a lot more important. Even if that means just listing that you contributed to some projects as an RA.

[Anonymous No.7334508]

>>7334499
oh ok I get you. thanks for the help man.

[Anonymous No.7334512]

>>7334484
Also, I'll make note of the fact that of my entire undergraduate graduating class (of which only a small portion were applying to graduate school), only myself and a small number of people that I had worked or was aquatinted with had been published (to my knowledge). It's also not as hard as you may think if you piggyback off of an already strong team.

[Anonymous No.7335724]

>>7334393
Functional analysis.

[Anonymous No.7335769]

>>7334450
>You should go to your library and pick up a copy of Artificial Intelligence: A Modern Approach and read it front to back.

Is it still recommended to read or should one wait for the new updated edition? About what kind of AI is the book? What if one is only interested in one type of AI? Can one just read the part one is interested in or does every chapter build on the next one.

Also what are the prerequisites math and cs wise?

[Anonymous No.7335806]

>>7334216
Here.
CS undergrad here currently working with NLP and machine learning. I am also going to apply to a graduate program this year.

[Anonymous No.7335963]

>>7335769
Methods for solving problems you might think are entirely disjoint tend to apply regardless.

[Anonymous No.7336190]

What do you guys think of the LSTM hype train?

[Anonymous No.7336212]

>interested in AI
>stuck in boring EE
>workload of EE doesn't leave me enough time to intensivly study AI

[Anonymous No.7336235]

>>7334393
>Also what are the mathematical basis of machine learning ? I heard statistics were essential ?
I'd say :
Statistics, Nonlinear optimizations and Functional Analysis.

[Anonymous No.7336299]

>>7335806
UCSC?

If not, what problem are you working on?

[Anonymous No.7336890]

watched some deep learning lectures...
they aren't fucking around with that there linear algebra and calculus

[Anonymous No.7337139]

Is it still worth getting into machine learning at this stage in the game?

I have no formal degree but enjoy programming and self-learning math.

[Anonymous No.7337232]

>>7334479
>>7335724
>>7336235
Thanks again. I'm surpised to hear functional analysis is useful, but that's good news since I happen to love the little I know of it.Nonlinear optimization is another matter though. Roughly speaking, what are the basis for it ? Is it a particular use of convex analysis (for instance) ?

[Anonymous No.7337237]

>>7334417
>obscure
Not really. We went over machine learning, albeit briefly, in a freshman class for computer science. I'm certain most computer science people know about it.

[Anonymous No.7337247]

>>7337139
You can just find a new model that make Deep learning BTFO and easily become a billionaire.

[Anonymous No.7337251]

I have a question, does any of you guys believe that Deep net is the end game? A lot of people is jumping into it now because computer hardware is better than 1970s.
What do you think is the weakness of deep nets?
I'm studying some Deep net models, but honestly I hope that it will meet a dead end soon and there will be something new, because currently training a large deep net is highly inefficient and it is only doable for big companies.

[Anonymous No.7337259]

>>7334216
I-I-Is it weird that I have gay feelings toward Andrew Ng after I studied his course on coursera?

[Anonymous No.7337283]

>>7335769
The latest edition (3rd) is reasonably new. I don't expect there to be another anytime soon, but I could be wrong.

From the Preface:

>The only prerequisite is familiarity with basic concepts of computer science (algorithms, data structures, complexity) at a sophomore level. Freshman calculus and linear algebra are useful for some of the topics; the required mathematical background is supplied in Appendix A.

I'm sure you can get something out of reading it non-linear. I've done that in the past. It's pretty good about referencing itself where the topics are highly connected.

It's a broad coverage and sometimes called "the agent book". Learning makes up only part of it. It also covers problem solving, knowledge, reasoning, uncertainty, sensing...

[Anonymous No.7337287]

>>7334216
Currently majoring in Computer Engineering. Are ML grad schools going to give me a weird look when I apply?

I'm considering switching to Softy, but I haven't done anything with circuits yet so I'm going to do one more semester of it

[Anonymous No.7337289]

>>7337287
If you are familiar with optimization problems, statistic and linear algebra then you are pretty much fine.

[Anonymous No.7337302]

>>7337259
tippity top kek

I made this austistc piece of work for you

[Anonymous No.7337318]

>>7337302
Nice, I think I will send this to him.

[Anonymous No.7337405]

>/sci/
>ML
pick one.

[Anonymous No.7337407]

>>7337405
easy there big guy

[Anonymous No.7337413]

>>7337407
4u?

[Anonymous No.7337583]

>>7334216
I'm a pure mathfag and I've dabbled a bit in the field. It's one of those fields where the moment you grasp the fundamentals you're suddenly struck by a profound sense that it's very important and has a staggering number of possible applications.

I've read that in regards to clustering algorithms there are people who study the topology of the clusters in your data space. Are there any books or references any of you would recommend for this?

>>7337405
Get with the program, anon. Machine learning is the future.
https://www.youtube.com/watch?v=AY4ajbu_G3k

[Anonymous No.7337592]

>>7337259
I took that course too the very first time they offered it (this was before it was called coursera and before there were other MOOCs around). I remember that as the course progressed he started looking more and more run down and didn't sound as enthusiastic while teaching. I initially assumed it was because he was bored of it and didn't really want to keep teaching us. Then, after the course ended, he put up a closing video where he thanked everyone, said he was proud of them, and so on. The guy sounded so sincere that I felt shitty for having judged him like that. Looking around the comments it looked like other people felt the same way. Setting up that course and doing all that shit was probably a lot of work and probably the reason he looked so run down all the time.

Andrew Ng is awesome.

>>7337302
saved

[Anonymous No.7337738]

>>7337251
Deep learning is the future mate. Waiting until quantum computer becomes popular and you will see real time deep learning everywhere.

[Anonymous No.7337753]

>>7334216
http://googleresearch.blogspot.nl/2015/06/inceptionism-going-deeper-into-neural.html

[Anonymous No.7337759]

>>7337753
top post off /r/ML
alrighty

[Anonymous No.7337777]

>>7337753
i wonder what researchers think about this sort of thing

http://arxiv.org/pdf/1412.1897v4.pdf
https://www.youtube.com/watch?v=M2IebCN9Ht4

also i think it's cool that these guys made a video for their research paper, more research groups should do that kind of stuff

[Anonymous No.7337888]

Do we have a way to uncensor Japanese porn yet?

[Anonymous No.7337894]

>>7337888
Yes triple eight. In Photoshop just look under Filters / Law Enforcement / BS / Enhance.

[Anonymous No.7337896]

>>7337894
I laughed

[Anonymous No.7337902]

>>7337232
>Roughly speaking, what are the basis for it ?
You have measurements, usually many of them so the system is overdetermined.
You try and fit those measurements to your model, you want to find the values of the parameters of your model basically
The equations you get are generally nonlinear and overdetermined. So yeah, nonlinear optimizations. But this is more like a small patch of the field. I'd just stick with statistics and FA, since nonlinear optimizations don't go much deeper than the žNewtons method.

[Anonymous No.7338036]

>>7335806
Im not from the usa. Im from mexico, currently working in sentiment analisys and authorship profiling.

[Anonymous No.7339118]

Thread is so fucking dead. Does /sci/ hate ML or are ML people really that rich and busy in the market?

[Anonymous No.7339219]

>>7339118
Thread topic is really general, and I'm guessing yeah most people [myself included] don't have the understanding necessary to discuss different models

[Anonymous No.7339829]

Idiot here. Can someone explain machine learning to me?

From the few articles I read so far it seems to be used for image/text/audio recognition. You train the machine by presenting it with input and then letting it produce an output until it comes close to the desired output right? Or is that just one way to do it? So the machine gets trained for one specific task?

[Anonymous No.7339850]

>>7339829
>You train the machine by presenting it with input and then letting it produce an output until it comes close to the desired output right
that's a recurrent neural network
http://karpathy.github.io/2015/05/21/rnn-effectiveness/

that blog has some good stuff, also check out this one
http://colah.github.io/
"neural networks, manifolds, and topology" is a good post from there for an actual explanation of how neural networks do what they do

but machine learning is more than just neural networks, that's just what a lot of people are interested in

[Anonymous No.7339851]

>>7339850
oh sorry i didn't read that line very closesly that's actually not a recurrent neural network deal

still check out those links though they're interesting

[Anonymous No.7339872]

Should I even bother reading tutorials on machine learning if I don't know any advanced math?

[Anonymous No.7339913]

>>7339850
I'm not that guy but those are pretty cool sites anon. Thanks.

I like that the groups page on Colah's blog uses Nathan Carter's approach to group theory.

[Anonymous No.7339935]

>>7339872
if you're just looking to make some basic machine learning programs you don't really need to understand the mathematics behind it

CS is all about abstraction after all, there are plenty of layman-friendly machine learning libraries out there

[Anonymous No.7339944]

I am interested in AI but also robotics. How do I choose what to focus on? To make things more complicted those fields by themselves are quite diverse. Making an AI sounds interesting but it's just code. Making a robot sounds exciting but without AI it's just a an empty shell. It seems impossible to do both in depth. Or am I wrong?

[Anonymous No.7339965]

>>7339944
Robotics is mostly machine learning nowadays, go into EE if you want to build something.

Then again machine learning is mostly "statistics in CS" with a fancy name.

[Anonymous No.7339970]

>>7339965
>Robotics is mostly machine learning nowadays
no it's not

[Anonymous No.7339984]

I am so high right now.

[Anonymous No.7339990]

>feeding your neural networks acid data

[Anonymous No.7340013]

>>7339990
Did they publish the tool they used for this? Actually, did the publish more than the blog post?

[Anonymous No.7340423]

>>7340013
There's a link to a small gallery in the blog post. They did not publish "the tool" because they're actually generated by messing directly with their trained networks. Such a tool is probably non-trivial to use.

[Anonymous No.7340460]

>>7339990
I would absolutely love to see a video fed through this network.