Deep learning thread. What is anon working on?

http://alpha.openai.com/miniwob/index.html

I wonder if anyone working on machine translators with neural nets on big scale
would be nice someday to forget about language learning completely

>>58042041
I'm taking Stanfords machine learning course on Coursera. After that I wan't to fuck around with OpenAI's Universe and try some challenges on Kaggle.

>>58042514
Nice, I'm going to be starting eDX's mini masters program for AI in a couple of months.

iTunesU also has some good classes on ML, can't remember the name of the proff, but he's a middle eastern dude and he explains things really well...

But with eDX i'll get a specialized, accredited MSc that can be transferred for credits if I wanta to finish my masters.

>>58042041
how can I make money with SEO and deep-learning
pls
teach me

can I make an advanced neural network mai waifu?

>>58042041
>deep memeing

>>58042289

AGI environments, right now ... genuinely interesting.

>>58042422
Google is. They are on par with humans as of the last few months. Already rolled out for 7 languages, the rest is following shortly.

>>58042041
What is the easiest way to get started with image classification? Is there anything that will run, and is easy to set up in Windows?

I have a folder of random images I'd like to classify (ideally with a net trained on ImageNet to get the most classes).

>>58044054
Probably tensorflow

>>58044188
Thanks! How long would do you reckon someone new to all this would spend setting it all up?

>>58044280
could probably be done in a few lines of code in python with tensorflow. training could take some hours, but you will probably not get it to work for a while if you are totally new to it.

>>58044323
I've installed python on windows before, and run some .py files. Do you just install python and the tensorflow python api and and pass a few lines of code to the interpreter? Let's say I could get a pre-trained net, could I just "load" it, point to my directory of images and have it chew through them?

Why isn't there any good, detailed step by step tutorials on this? Taking the user through every step from start to finish?

So it this anything more than statistic?

>>58044443
It's literally hundreds of tutorials on the web.

>>58042041
What have you actually implemented so far? I always got the feel that deep/machine learning was mostly about reading 300 books in order to create a retarded AI with 20 lines of code

>>58044668
Can you show me a good one? Not a single video tutorial on youtube for instance.

>>58044561
Is your brain anything more than statistics?

>>58044826
Of course not, statistics are racist and make non-asians look bad.

>>58044561
That's a dumb question.

File: blondeQtpie.jpg (147KB, 1080x1080px) Image search: [Google]

147KB, 1080x1080px

>>58044826
>>58044955
We had this deep meme thread few days ago with 300+ replies and several people admited that deep learning is just google buzz meme.

Its just associated learning

Not a breakthrough

Don't know shit about deep learning, but I'm developing a tool that uses bayesian inference to estimate parameters of censored distributions

>>58042041
I'm pretty sure I have zero-order and first-order logic down. Working on second-order and higher-order logic.

To learn deep learning get a background in:

Calculus I-III
Linear Algebra
Statistics
Probability
Optimizations

Just using tensorflow and libraries is not enough

How do they even determine how each node should handle input?

>>58044443
>on windows
Expect trouble
Quadruply so if you intend to use CUDA to speed it up.

>>58049518
I have an AMD card. Am I fucked?

>>58049705
It means you will not be able to use nvidia's biggest selling point to speed up calculations.
If you don't care about speed, it's no big deal.

If you do care about speed, it may be worth trying some other library, like theano. Though all of them are clunky on windows, because windows is an environment to run video games, not neural networks.

>>58049785
Are there no OpenCL support or other dirty tricks to run CUDA on AMD?

>>58049935
You can't directly translate CUDA into OpenCL.
Then there is a whole set of precompiled nvidia libraries, just for neural networks to go fast.

With an AMD card, you can either find a library that allows OpenCL (Theano does that) or deal with no GPU support.

>>58049982
It's the training that takes the most compute time though, right?

>>58050064
Yes.

Unless you want to do some recursive calculations, like DeepDream shit. Those will also take a lot of time.

>>58050097
What about this branch of caffe?

https://github.com/BVLC/caffe/tree/opencl

>>58050232
Could work.

But one important thing to note:
Follow the tutorial EXACTLY. If they say you nee a version 1.2.3.2.1.3 of dicksmasherlib and you have 1.2.3.2.1.4, downgrade. Unless they say like "x or newer", then it should be OK (but it's still better to get the exact).

I once tried to set up one of those libraries - on ubuntu, where it is much easier than on windows - and wasted days on trying to get it to work with a slightly newer library.

>>58050339
Good point, thanks!

Do you have a working setup?

Would you be interested in setting up a working linux VM and sending me the vhd in return for some bitcoin for your efforts?

Basically make everything ready and type out the lines I need to load a pre trained model, load my source images and spit out the classifications data. Only needs to run on CPU, classification is usually less than 100ms/image on CPU, right?

>>58050520
>Do you have a working setup?
Not really, since I had to reinstall and wipe the HD.

Doing it on a VM would be quite hard, since most VMs have limitations on GPU access. I think the dude I was setting it up with (it was n university project) wanted to use a VM and gave up, because none of them supported the necessary kind of access to GPU.
We worked with CUDA, OpenCL would be even harder.

The best way would be if you dual booted it.
This is safe and has complete tutorials. Scrap some HD space (can be old HD you aren't using or something like that) for a stable ubuntu and install it.

If you can't, first set up a CPU-only setup, then progress to GPU one if you realize you need it.
CPU-only should be easy to set up. If you have a good CPU, the GPU support may not be necessary, though it is always good to have.

>>58050635
I have dual Xeon E5-2670, and my machine can run 24/7. So I only need CPU.

How do I convince my friend not to do a masters in ML?

>>58051396
Force him to try to set up the environment for them, then watch him rip his own dick out in frustration.

If he passes that, there is nothing you can do, because ML is the shit and more than just a meme.

>>58044026
>translators BTFO
>chess players BTFO
>go players BTFO
>medical diagnosers BTFO
who's next?

>>58042289
So.... what's the purpose of this?

Reminder that reservoir computing is the future

>>58051519
Truck drivers

>>58049127
Are there any other math course aside from those that can come in handy here?

My standard CS course covers everything except Calc III and Probabilty, but I can get those by simply taking mathematics as my mandatory elective subject. If I do that, I'll need another one or two extra math courses to complete the elective.

>>58051519
Once they replace programmers, singularity becomes something more than just a meme.

Can someone explain what kind of problems these deep learning networks are good at and what kind of problems they're not good at/capable of solving?

>>58052884
Good:
>natural, animal stuff - image recognition, speech, stateless processing with fixed input size
>art
>pretending to be an unintelligent human or a woman

Bad:
>logical, mathematical stuff - pathfinding, routing, designing
>retaining state for too long - actually understanding an article longer than one sentence

Neural networks in general are more of artificial intuition than artificial intelligence. When they know an answer, they still don't know why.

Basically, if a woman can understand something, a neural network will get a hang of it. If you need a man, a neural network will fail miserably.

>>58042514
I've just started it. I feel like I should have brushed up on some statistics first as it's been years since I've used any seriously, but it's going OK, feels great actually putting statistics to use.

>>58053074
That seems like an incredibly superficial analysis.

>>58053368
Didn't know you wanted an autistic in-depth one.
If you want one, it's a great idea to get out of /g/ for a longer while and find a multi page article about it outside.

>>58053197
Also saw the AI class on coursera but I was too poor this month to sign up

I want to start in January (assuming it's offered again)

Is it worth it? How deep does the teacher go?

File: 1479533373583.png (244KB, 500x524px) Image search: [Google]

244KB, 500x524px

Just finished a classification NN from scratch, last week. Feels good to finally understand how it's done

>>58053668
The course is free, you only need to pay if you want the certification. About the course, it will teach you the basics and hand guide you all the way to the end. For someone who has never seem or read about neural networks and machine learning before, the course if obviously not enough to fully understand what you've been taught. I highly recommend trying to implement what you've learned by yourself, in your prefered language, after you are done with the course

>>58044561
>>58045431
the only thing that's "deep" about it is the structure of the neural network. People buzzworded it up and think the "deep" means like it has intuition and dreams in code and shit.

>>58054649
The learning part may also by "deep", if the algorithms used are adapted to hierarchical structure, or at least big depth.

Say, a vidya gaem playing network being taught to recognize objects and make decisions separately, then combined.

>>58054649
Who ever thought "deep" in deep learning referred to anything other than the amount of layers?

>>58053074
(You) are the sole reason why feminism exists.

>>58055722
Then (((I))) will be the one to destroy it.