Machine Learning / Deep Learning General

>>57762755
No one is smart enough to do machine learning here

>>57762755
We really should have a data science general though. To cover more topics and concepts.

How are you going to do any meaningful machine learning with enough DATA and PROCESSING POWER?

Both aren't really available for consumers.

>>57762876
That's what I'm trying to do

>>57762891
*without

>>57762861
I'm turning in my multithreaded neural network as a term project this friday lol

>>57762958
Undergrad or Grad student?

Also, what are the best grad schools for deep learning besides Washington?

>>57762992
CMU

I do machine learning but bioinformatics. Grad student here UCSD

File: chopsticks.jpg (725KB, 1022x1318px) Image search: [Google]

725KB, 1022x1318px

>>57762992
underggrgrgrgrggrgrgrgrgrgrgrgrgrgrggrrgrrgrgrgrgrgrgrgrgrgrggrgrgrgrgrgrgrgrgggggggrrrad

>>57763032
I want to get a list of schools to apply to.

So far I have,

Washington
CMU

That's about all I know even though I know there are far more.

I'm interested in Neural Nets. I come from a mathematics background, but recently started beefing up on core CS courses.

File: 1478198641179.jpg (28KB, 500x555px) Image search: [Google]

28KB, 500x555px

>>57762755
I recommend Python for Machine Learning for anyone trying to break into the field and get an exposure to what's out there.

Also Andrew Ng's coursera course followed by the recorded course he did at Stanford.

>>57763067
I meant to add I'm mostly interested in image recognition problems. What are the big problems in this area currently?

File: 59d4574e03a0bd27958bbaf7e5edf66a.jpg (135KB, 750x1000px) Image search: [Google]

135KB, 750x1000px

>>57763111
trips of why don't you figure it the fuck out yourself

>>57763125
because im fairly new and was curious what the open problems were. I'll just read papers and find out as you suggested thought maybe there would be someone that researched this on here

File: 02f7a1ee7a964485bf829bab69a8c74c48b5a0ce.jpg (319KB, 1200x1600px) Image search: [Google]

319KB, 1200x1600px

>>57763146
I think you should use this thread for its intended purpose; I just wanted an excuse to post Asian cleavage

>>57763067

Apply to SFU in northwest BC Canada. We have a big data program.

>I come from a mathematics background

Good, because a lot of people who get into it don't know what a partial derivative is.

File: math.jpg (9KB, 275x183px) Image search: [Google]

9KB, 275x183px

>>57763168
Thanks for the pro-tip. Can't believe they wouldn't know what a partial derivative is... hard to not know what that is and fully understand gradient descent kek

>>57763190

My background is physics, so the math comes easy. If you can implement numerical methods by scratch, you're a god among the rest.

>>57763159
seamonkeys don't count.

File: Brown Representability.png (1MB, 1344x1196px) Image search: [Google]

1MB, 1344x1196px

>>57762755
This is a must read:
http://www.deeplearningbook.org/

https://www.youtube.com/playlist?list=PLbuogVdPnkCozRSsdueVwX7CF9N4QWL0B

>>57762861
>No one is smart enough to do machine learning here
Aren't you overestimating ml?

>>57763325
what is it? prepackaged in libraries for anyone to use now?

File: Iterative_Places205-GoogLeNet_4-676x450.jpg (142KB, 676x450px) Image search: [Google]

142KB, 676x450px

Somebody had posted this awhile back on /g/.
https://github.com/ryanjay0/miles-deep
On an unrelated note, are there any programs where I can train models to create images of real world objects? Something like deep dream, but not psychotic?

whats the difference between machine learning and deep learning?

https://github.com/josephmisiti/awesome-machine-learning

https://github.com/ChristosChristofidis/awesome-deep-learning


some pajeet who does machine learning videos

https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A/videos

>>57763493
Let X>>Y denote "Y is subset of X"

Then I see it this way

Artificial Intelligence >> Machine Learning >> Deep Learning

>>57763505
>https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A/videos
Wow, based videos. Honestly, guy is super dope. Thanks.

>>57763493

Machine Learning is an umbrella term for data analysis techniques. Regression, neural networks, clustering etc, fall under machine learning.

Deep learning is an extension of neural networks. It's basically neural networks with much more layers used. For example, see convolutional neural networks (LeNet and GoogLeNet).

>>57763380

Keras and TensorFlow greatly reduce the work needed to get a NN running. Specify your data and optimization method and you're good to go.

>>57763621
if its that easy then why do companies pay top $$$ ML engineers?

>>57763637

1. ML is hot and companies don't know shit about it, so they're just throwing money at it.

2. You need to understand the underlying process to tweak the parameters, know what method to use for what situation, etc. There's a bit more to it than just throwing your data into the program and expecting good results.

The libraries are indeed getting simpler to use. No need to implement something from scratch, unless the method isn't included in the libraries of course.

>>57763674
then ive been doing it wrong. i've been implementing all my shit from scratch to understand all the underlying mathematics of it

>>57763686

That's good though. Don't stop doing that.

>>57763702
Ok cool. Thanks for the pro-tips

>>57763686
No, that's the right way to do things. Once you understand what's going on under the hood, you can make better use of the libraries.

Why are most ML researchers (specifically François Chollet) insufferable libtards?

>>57763754
what are his political views?

>>57763942
See for yourself.

https://twitter.com/fchollet/status/751531675113299968
https://twitter.com/fchollet/status/755272039150211072
https://twitter.com/fchollet/status/796292602542723072
https://twitter.com/fchollet/status/803797013124894720

I could keep going....

>>57763754
Because conservatives as the name says are about conserving. They don't care about technological progress. Progressives do. It's in the name. All technology is inherently liberal.

>>57763953
He is insufferable. What's his educational background? LinkedIn doesn't show

>>57763962
Then why is financial industry as a whole much more conservative despite being at the forefront of technological progress?

>>57763989
Because it's been co-opted and corrupted. Fintech isn't developed by in the financial industry. Also, please note that "liberal" in rest of of the world means "libertarian" in the US.

how can I use ML for making my 2D games better?

>>57762876
We really shouldn't

>>57763962
>All technology is inherently liberal.
Behold fully automated luxury gay space communism.

>>57764090
Ai for enemies?

It would take alot of work to make a neural network play a complicated game.

>>57764237
>>57764090

While it may take a bit more work than a simple AI, an agent trained through Policy Gradients w/ Deep Convolutional Networks can be done without too much work nowadays. The main issues are:

1. AI movement will likely appear very sporadic. Reinforcement learning doesn't generally penalize moves that look "weird" but don't have any actual effect. For instance in pong the paddle instead of remaining still while waiting for the ball the start bouncing back may jitter a ton. This doesn't look human, but it doesn't actually make the AI perform any worse so it's penalized (at least generally in the literature, you could probably experiment with "regularizing" the network to penalize extra movements slightly versus staying still).
2. It takes a very long time to train well. A friend of mine wanted to create an AI for a game at a 2 day Hackathon and when he asked me about using Machine Learning for it, I told him that he probably wouldn't have enough time for it to finish learning. Expect even with a GPU for these methods to take days to get good on any sort of moderately complex game.

File: 1455391820078.png (15KB, 785x924px) Image search: [Google]

15KB, 785x924px

Are neural networks and machine learning a meme?

What's the best card for DL right now?

Has anyone trained google's deepdream stuff on pictures of cocks?

I'm currently working on a furry porn recommendation deep learning network, based on a truncated inception network.

All these autistic furries accurately tagging their porn gives me a somewhat decent dataset to work with.

>>57765061
Yup.
https://open_nsfw.gitlab.io/

>>57762755
>>>/x/

>>57763754
François Chollet doesn't even know the first thing about ML. He's such an inbred holy shit. His library is beyond unusable, it's beyond pathetic. It's a true wonder and miracle that anyone at ALL used it since it's literally not usable. What's impossible to understand is the fact it gained traction. I mean the thing isn't even remotely stable, let alone the fact it has more bugs than a bug collection wrapped into a software, and less features (mostly but not entirely due to being lcd between tensorflow and theano) than raw cudnn.

>>57762861
yeah? how bout you use this thread to educate them rather than acting like a pretentious cunt

>>57763953
>Feeling jealous of the version of me who is waking up in the right timeline, while I will be forever stuck in the dystopian fork of reality.
well he is not wrong

>>57767658
That's impossible. In the first few threads, /x/fags made sure to establish that any and all competent discussion is disallowed, chasing out the only 2 people on /g/ who know anything about ML.

>>57763621
Nobody who knows even the first thing about ML uses keras nor tensorflow.

>>57767713
>im fucking l33t hacker
>oh yeah? sounds interesting you should explain more
>no i cant 2 many fags in ths thread so i cant actually explain it to u but i DO know how 2 hack xd

>>57767764
Thanks for confirming that the ban on all ML discussion is still in effect in this thread. Wouldn't want to post seriously in an /x/ thread.

>>57767720
What are hey using? What reasons do they not use TF? How is theano?

>>57767720
This

>>57767782
Protip: they write their own libraries, that's why they're getting paid

File: Screenshot from 2016-11-30 08-29-19.png (70KB, 1360x768px) Image search: [Google]

70KB, 1360x768px

>>57767779
>Wouldn't want to post seriously in an /x/ thread.
alright then get the fuck out, useless faggot.

I've been working on a box 2d editor for a while and soon I'm going to be rigging some ragdolls up in C++ and trying my hand at machine learning.

Here's a cool biped neural net in 3d: https://www.youtube.com/watch?v=pgaEE27nsQw

I still have some work to do before I'm satisfied with the editor but I'm planning on rigging the ragdoll's joints up to a neural net. I haven't done too much research into it but it sounds like there are many prebuilt libraries. Set the goals I want, and reward the neural net for behaving in the way I want until the desired result. I'm thinking of training it to do things like staying upright and walking

>>57767852
Deeplearning4J comes to mind when you say that.

File: 1458228073143.png (464KB, 867x761px) Image search: [Google]

464KB, 867x761px

>>57767872
Noice, anon. I would like to also start playing with neural networks, would you be kind enough to post an update (or more) when you are done (or at least made progress) ? You can make a general and post it.

thanks and godspeed with your work

>>57762861
It's nothing more then a linked list mi nigga.

>>57768111
Yeah for sure, hopefully I'll be able to finish this damn editor in a day or two. I've been working on it for like a month or more and I'm almost done. If this general is still up I'll get back in it and post.

>>57768111
here's the editor by the way it's kinda fun to play with
https://krogank9.github.io/box2d_editor/editor.html

>>57762891
digitalocean.com

>>57767872
What music are you listening anon

>>57768264
some vaporwave

https://www.youtube.com/watch?v=daFXlcK_8Js
https://www.youtube.com/watch?v=_8UGxPxYZS4
https://www.youtube.com/watch?v=Z1Z0KByg-KM

>>57768300
DIfferent anon but love it thanks

>>57762992
>>57763032
>>57763067

Wait, why Washington?

My thoughts on the top three in the US would be:

Stanford
NYU
CMU

>>57768738
Washington has a goat informatics faculty

>>57769056
Not quite deep learning though

>>57762755
Is Linux necessary to study machine learning?

>>57769244
Not necessary but you should use it anyway.

OTOH tensorflow r0.12 supports windows finally

>>57768738
Washington has a very large ML research faculty

>>57762755
"Galactic Leyline".

G A L A C T I C

L

E

Y

L

I

N

E

ok someone enlighten me...

I know shitall about this

Why should I learn it?

About me, social scientist with advanced statistical knowledge. So I can have a shittone of data in a blink of an eye if needed, however I learned R to cope with the SPSS bullshit to speed up processes for my personal research.

Can Machine learning help with that or make me rich in any way as a data-analyst?

>>57770367
Machine Learning is hot, but if you pass yourself as a data scientist then you'll be very in-demand on the market. You said you use R and have advance statistical knowledge.... then market yourself as a data scientist NOW.

You are very employable...

Deep learning is a very niched area of machine learning and not all companies are using it/need to use it. But applied statistical knowledge is so hot right now. You could and should apply to be a data scientist.

I recommend FORTH or APL for machine learning

File: 1469276878927.jpg (81KB, 720x720px) Image search: [Google]

81KB, 720x720px

AI waifus when?

>>57767720
>people actually believe this
Next you'll say no one uses Caffe, Torch, or Theano either.

>>57771753
Now I claim
Outlaw star cutie

>>57768181
Does digitalocean even have GPU instances? I thought most people would use AWS or Google Cloud Compute/Microsoft Azure

File: 1477798823197.gif (4MB, 450x253px) Image search: [Google]

4MB, 450x253px

life support bump!

>>57767782
>What are hey using?
Theano for low level with lasagne for high level in python, torch in lua, or caffe in C++ (although caffe has never been particularly popular and it's even less popular nowadays, probably because scientists don't like manual memory management and associated bugs - in particular, scientists are not software engineers).
I've also heard a few good things about mxnet but that could be because I was working closely with people who had helped in its development.

>What reasons do they not use TF?
TF has only 2 advantages over theano:
1- faster compile times (doesn't apply to rnns because TF can't have compute loops in the graph so it unrolls instead, which is orders of magnitude slower than theano - and only in rnns do compile times matter, so it's a fake advantage)
2- tensorboard (but the visualization is mostly bloat that means nothing unless you have engineers and PR people in the loop, or when running at google scale)

Theano supports way many more operations and is significantly faster at runtime (part of the reason is that the public release of TF is not the same thing google uses internally, for instance they use a custom nvcc, and when they first released TF publicly, this caused massive breakage where some models were literally slower on GPUs than CPUs because of that - however a large part of it is simply that theano has significantly more optimizations than TF).

>How is theano?
Great. Unlike TF, it's actually possible to debug a theano compute graph and a theano model. For instance, try to figure out which ops are running on GPU vs CPU in TF. It's impossible. In theano? Set a flag and you have the answer. Want a printout of your theano graph as a graphviz graph? Sure, there's a flag for that. Want to selectively enable or disable optimizations? No problem. Want to display a text representation of the graph? Just say so. TF has fuckall, or pure noise, by comparison.
It also supports a fuckload more operations (cont)

>>57768738
NYU is probably the best right now, stanford is actually dogshit for ML (most of their results are falsified or misrepresented, but they also sometimes turn up good shit). CMU and MIT are good as well.

>>57776732
Most importantly, theano actually supports loops (scan) and vector modifications (set/inc_subtensor), correct complex broadcasting, scalar-to-tensor max/min/etc. and other pretty damned important features.

File: [EG]Outlaw_Star_02_BD(1080p_10bit_DualAudio)[2B7F131C].mkv_snapshot_21.34_[2015.02.04_22.47.53].jpg (360KB, 1440x1080px) Image search: [Google]

360KB, 1440x1080px

>>57762755
>Not pasting remastered Melfina

>>57776889
next time can we please use an image that isn't a ugly as fuck

>>57771753
not any time soon.

>>57776840
Forgot to mention: theano tends to have bugs every other version. Typically you won't encounter them, and if you encounter them, they typically get fixed fast and can easily be bypassed. The same can't be said about TF, and most certainly not about keras which has years-old bugs that chollet decided are 'user error'. I always find looking at the issues funny because chollet is the most thumbs-down'd/disapproved of person in any github issue since he's always wrong about every single thing he says, as if he doesn't even know what his library does or can do, but I digress.

Suffice is to say that in python land, theano is the best option, but being so low-level, you want to use lasagne on top of it (almost exclusively for optimizers and regularizers, though, because the rest isn't any issue to implement in raw theano - but hey, might as well use the whole stack).
If you want to do ML in production scale, I would recommend torch instead because it's easily embedded in C/C++, is much faster, and most importantly, it has no GIL (the company I work for had to write a custom C++ runtime that takes the saved theano graph and runs it with rewritten ops because the GIL was fucking shit up when running a model as a web service).
Caffe is worth a look at if you're more adventurous, but it's a lot less used and certainly much harder to use (both torch and theano/theano-based stuff can do automatic differentiation or symbolic differentiation, the same is not true of caffe last I heard, among other things). At the rate TF development is going, I don't expect it to be usable for the next 5 years.

>>57762755

Goddammit that TensorFlow presentation video made me cringe hard.
Why are all these presentation filled with tons of buzzwords and hipsters, just fucking write the program, chances are it'll run faster.

>>57777242
It's pure PR. Have you seen the IBM Watson presentation? 35 minutes of pure, unadulterated lies spoken as pure buzzwords. Whereas IBM Watson targets clueless heads of companies, tensorflow targets clueless 10 year olds.

>>57763403
Where can I download the training data son?

>>57762755
Just as I'm trying to get into this, here you are right on cue.

I'm wondering how to implement machine learning into what I'm doing. Right now I'm only programming a rules-based Python chatbot which uses web scraping to get information on a section of law, as well as policy and industry definitions for the user. This model is far too limited though.

From the limited research I've done it looks like I could use pywit (wit.ai Python module) for the 'conversational' elements with web scraping functioning alongside it. Then the next step would be to look into using a larger framework such as NLTK to do away with the need for web scraping. I could be completely wrong, but that's my hypothesis right now.

In any case, it seems like a really exciting field to get into (even just viewing it as a hobby, I'd love to create AI) and I'll be looking through all the resources in this thread with serious interest.

>>57776758
Agreed. I haven't followed too much of MIT, but Stanford keeps running into the problem of all its best professors leaving for industry. Could you elaborate more on the Stanford/results stuff

CMU used to be good too until Uber poached its whole team too.

That leaves NYU, with Yann holding the fort since Facebook AI is next door so he doesn't need to leave.

File: ml-sp.png (26KB, 528x514px) Image search: [Google]

26KB, 528x514px

>>57762861
>No one is smart enough to do machine learning here
Speak for yourself.

>>57776951
TF has dynamic unrolling now.

>>57776840
>>57776732
>>57776951

Thanks for the response. Your response backs up what one of the professors I spoke to said. They recommended I use theano.

I have an Radeon AMD 455 GPU. Can I use theano/the radeon GPU? Most searches I see come up with NVIDA/CUDA or w.e.

Also, what schools should I put on my map for graduate studies in deep learning?

Based on the comments here NYU just got added to my list.

>>57776951
Curious - what is your background?

how do you make money off ML?

>>57780038
Nobody say finance.

NOBODY SAY FINANCE.

>>57780038
Predict stock market :^)

File: 1c3f7433f2d7e945cb4108c31c58f85cb8d2b89e722ddb7634a90844efe2dc9c.jpg (80KB, 500x497px) Image search: [Google]

80KB, 500x497px

>2017 - 1 month
>still not using maxout activation functions
GET WITH THE TIMES YOU PLEBS

Maxout is probably the coolest deep learning idea that I've seen

>>57780357
>at least twice the computation
>minimum performance gain

>>57780038
ALL IN JNUG AND YANG

>>57780372
>using shitty libraries not designed for maxout
>not rolling your own optimized implementation because of how simple it is

>>57780332
I actually want to try to use ML to predict stocks. I am able to get data from this API http://financequotes-api.com/

When my semester is done I'm gonna try to work on this as a project. Do you guys know if I can use tensorflow in Java or if there is an easy way t link Python with Java.

testing to see if I'm still banned

>>57780497
you're still banned on /r/ml

>>57780488
People on wall street with beefy machines and proprietary data are gonna eat you alive

Did a fun project this semester where I used Generative Adversarial Networks to generate human faces. It eventually crossed uncanny valley which was exciting.

>>57778324
I know TF is planning to have a scan op but as far as I know it's not live yet. The important part is that no unrolling at all is permissible in non-trivial RNNs. Sure it takes 30 minutes to compile a theano graph under these conditions, but I'd take that any time over the 11 hours it takes tensorflow to compile the same graph (those are real numbers, using keras to compare models).

>>57778942
Nothing serious supports opencl right now, it's all cuda (primarily because cudnn is so good). However there's experimental opencl support for theano somewhere and there's a branch of torch that has opencl support. Your safest bet is most likely torch-opencl as it's probably more mature.

For DL, the best option by far is UdeM. In the first place, Yoshua Bengio is basically the only professor with a solid DL background that still hasn't gone to industry. In the US, CMU, NYU, MIT and Berkeley, with washington and purdue (to a much lesser extent) as last options, are possible options. Washington is a good choice, but less so than CMU, NYU, MIT and Berkeley, by volume of meaningful ML research output.

>>57779928
CS undergrad, CS (ML) master's (DL, mostly images, mostly generation), 3 years as a pure research scientist in industry (mostly dialogue, nlu, some mrc, some nlg).

>>57780038
Go to industry or target a domain and make a startup. Then there are obvious "you'll never make it" targets like >>57780332

>>57780758
use VAEs next time :^)

>>57780795
a fork of torch*

>>57780728
I'm not gonna actually bet on the predictions. It's just a project.

>>57780795
tf.nn.dynamic_rnn

>>57780795
What's MRC?

>>57780926
machine reading comprehension

How can I help Trump's presidency with ML?

>>57780795
Cool, where'd you do your masters? That's kind of what I'm applying for now currently.

>>57781136
NYU, under lecun.

>>57781201
Sweeet. That's one of the places I'm gunning for. Is he still active at NYU?

What's the best method for regulating RNN?

>>57781211
Lecun said:
>Will this change your relationship with NYU, Courant and the Center for Data Science?
>I am going to step down from being the director of the Center for Data Science, now that Raghu Varadhan has been appointed interim director. I very much enjoyed creating it and getting it on track. I will still teach, I will keep my lab with my students and postdocs. I will keep a research activity here. I’m probably going to change the focus of the research in my lab to focus more on research and principles, and less on applications.
In an interview. I haven't checked in person or anything so I don't know how active he is, or how much research interests have changed, though.

>>57781288
*regularizing

>>57781288
zoneout or the version of recurrent dropout that applies to gates. Hope you're not talking about vanilla RNNs though.

>>57781306
thanks

>>57780824
I've done some research and to my understanding, it's still unrolling the graph. The word 'dynamic' means that you may have different-length timesteps and the graph stops unrolling instead of continuing with e.g. 0-padding, not that the rnn isn't being unrolled. If you have stronger evidence of the contrary, I'd really like to know about it.

>>57780795
>>57781347
Why is unrolling not permissible in non-trivial RNNs? For skip-forward across time-steps you can just use a counter to track how many time-steps have passed.

>to read list now 500 papers long
help me /mlg/

>>57781530
The problem isn't in runtime but compile time. That graph that takes 11 hours to compile instead of 30 minutes means I can't tell whether the model is performing well or if I have to change some parameters, if I'm allowed to proceed to the next step or otherwise, for 10.5 hours longer. Not only does it mean I can run 22 (dependent) experiments in the same time in theano as I can in TF, it also means I can finish two weeks' work in a day provided all experiments come to fruition.

Of course, if you already have a properly trained model, you barely care that it takes even a full day to get the graph compiled, since only runtime matters afterward. TF is still slower than theano in that department, but with their (not so) recent fixes (nvidia nvcc compatibility and cudnnv4+, mostly), it's still viable.

So the reason unrolling is not permissible for non-trivial RNNs is not the unrolling itself but rather the compile time associated with the unrolling, and it is not permissible because the research loop very strongly depends on results of a previous batch of experiments to progress: get an idea, read, implement a model, run then tweak based on experimental results then go back to run until satisfied, then implement something based on the results of that loop, etc. for that idea. The run part requires recompiling a graph every time. Thus, unless the graph compiles in reasonable time, you don't get to do research.

>>57781717
..but there's almost no compile time with tf.nn.dynamic_rnn? I am using it right now.

>>57781061
You can't. His resources are already far better than you can ever hope to be. Just sit back and enjoy the ride.

>>57767720
this board is absolutely terrible

>>57781717
State size 2048, 128-long, 4-layer GRU with Layernorm, multiplicative integration with minibatch size 128 takes less than 15 sec to compile with tf.nn.dynamic_rnn(). Sure it's 20% slower than the tf.nn.rnn() version, but that's all it needs for experiments.

>>57781779
But I want to make SJWs on reddit and MSM squirm and beg for mercy

>>57781947
make a pepe generator

>>57780795
In that case is it perfectly fine to train my neural nets on the CPU?

I have an older computer with an NVIDA 750 M GPU. Could that take advantage of Theano GPU utilization?

I'm new to DL. I have a pure math background and recently gotten into it. Want to get a PhD in the field which is why I asked you the target schools, thanks for the list!

Do you think DL and/or ML will still be big in 6-7 years (by the time I have a PhD)?

can pajeets into machine learning

How many porn images do I have to feed my model before it can generate something actually satisfactory?

Can ML ever explain why Maki is so cute?

>>57782623
Something about this page just always cracked me up: http://waifu2x.udp.jp/

>>57782698
who's hyped for nips

everyone get hyped for nips

File: 1480428843638.jpg (111KB, 751x1000px) Image search: [Google]

111KB, 751x1000px

I'm a footfag. Can I still do ML?

>>57782818
No. You're beyond saving.

>>57781947
>But I want to make SJWs on reddit and MSM squirm and beg for mercy
I want to write a tool to determine authorship of posts so I can find alts of SJWs.

Where do I start? I'm downloading the corpus.

>>57782290
Got a lot of indians at my university and from past experience I can tell you : nope, most can't

File: tumblr_oh4tdni4m01t08m8eo1_1280.jpg (872KB, 1271x1920px) Image search: [Google]

872KB, 1271x1920px

>>57783013
:(

>>57783081
Kys faggot

Nice to see such a thread.
I've worked with one of my university's work groups on active machine learning topic for a while and am now doing my master thesis in that field.
Currently digging into SVM to ensure I got everything right there. Did anyone of you give the paper by Vapnik on SVM density estimation a read?

>>57783141
SVM is already obsolete

New guy here, been interested in this field but never found the time to experiment on my own. What are some cool projects I can work on? I'm motivated to teach myself but currently ignorant and unsure what I can pick up as a personal project.

I took a class years ago on Coursera and followed along and did the hand-written digit classification project. Thought it neat and kind of want to revisit it. Could make something that can read handwritten letters and document them.

I also remember a website that generated fake handwriting which was neat. Anyone know what I'm talking about?

>>57783172
If by SVM you mean Support Vector Machine, can you elaborate?

>>57783755
SVM only works better than deep neural networks when # of observations is small. You can get shit tons of data in almost any industry nowadays.

I have a class where we work with TensorFlow for Deep Learning/Machine Learning

In the Tutorial for Instance, we are working trough is this:
https://github.com/pkmital/CADL

But our professor says that it's not perfect.
For Instance Session 2 where they don't tell you that you should use a "One Hot Encoding" for a better Output and the output Pictures therefore look too similar.

Is there a point to learning ML if I can't afford expensive hardware? Like can I make side projects by myself that won't require me to bust my wallet

>>57784646
Learning ML is useful regardless of whether you have expensive hardware. Practical ML and especially DL on the other hand require some investment.

i'm not sure i can trust /g/ with something this high level

>>57780488
You can use Java to do deep learning projects. Look up deeplearning4J or skymind.

To the Java guy
https://deeplearning4j.org

On mobile so I can't spot your second comment.


To anyone else
What problems in image reciognition are "open" right now?

Where are good resources to learn/read about such problems?

Will deep learning still be hot on 6-7 years? About how long it'll take me to apply and finish a PhD. Still have a year or two before I apply

Question: What are good text-mining libraried for either Java, PHP or Ruby?
I need more or less basic stuff like text-highlighting/recognition, they only problem is that i need a library or algorythm that does not require learning.
Didn't know where to ask, /dpt/ aka /homeworkgeneral/ didn't seem like the right place.

>>57781815
>>57781735
It depends entirely on timestep size. Do you have at least 200 timesteps? If so and if it does compile that fast for you (don't believe the tf compile time, it seems to report almost immediate completion and then does the actual compilation for the 11 hours I mention), then you are right and it definitely doesn't do unrolling.

>>57781989
Not for non-trivial tasks. The speedup from using the GPU can easily be 100x+. Your 750M is a better target but you'll probably have to explicitly disable cudnn use since that typically requires later compute levels.

>>57784646
As >>57784685 says. RL, on the other hand, ends up working just fine on CPU, so that's an option as well if you just want to do sota AI of any kind.

>>57785237
Image recognition has been solved for like 5 years now.

>>57785674
128 timesteps. 15s is the time from tf start building the graph to start training.

>>57785704
OK, then this is good evidence that it doesn't unroll the graph. Good to know. 20% slower than rnn is just within the margin where it's still usable for experiments so this fixes that.

>>57782818
Do you have 1,000,000 images of foots?

File: tumblr_of886x3rne1t08m8eo1_1280.jpg (100KB, 637x960px) Image search: [Google]

100KB, 637x960px

>>57785736
I only save beautiful ones I have around 10,000. Does retraining/finetuning help?

>>57785674
Why's RL plausible on CPU?

>>57785822
Mиpocлaвa Цибyльcкaя

>>57785822
1. skew, rotate, translate to expand feet-set
2. autoencode into foot2vec
3. use GAN to generate infinite feet
4. Publish FootNet to arXiv, under anon et al
5. Put the entire paper in footnotes

>>57786256
RL is inherently sequential: you have a state, action, and reward signal. You need the next state and next action space (typically it's always the same, though) to compute the next reward signal. ML can be used inside RL (for function approximation, as opposed to the classical approach of maintaining a state action matrix), and ML can use RL (such as with RLNTMs or in large systems such as non-end-to-end dialogue systems), but RL itself is inherently sequential and thus does not benefit much from GPUs.

>>57786309
Is it not possible to parallelize an RL run using GPUs? Or is it mostly unsuitable for that task? (E.g. different runs would have different lengths, so parallelization doesn't really work)

Also, for the game-playing RLs, do they not use enough DL at each step to justify going to the GPU?

>>57786338
There is no way to parallelize a generic RL problem beside e.g. running multiple off-policy searches simultaneously in case of off-policy learning. In the case where the world state can be eventually reset, it can be viable to train multiple copies of the agent and then e.g. average the learned policies.

Even for sota game-playing, DL isn't necessary (obviously, it depends on the game). The DL part of RL, when used, is always run on GPUs, but that's for function approximation (e.g. the value network or the policy network depending on algorithm), not for the actual RL proper.

>>57786420
I guess I was thinking in terms of convnets for processing the images of games.

I wish that Deep RL were a little better documented, though I guess it is a little newer to the scene compared to CNNs/RNNs.

so i'm literally have zero knowledge on machine learning and wanted to give it a shot.
i am reading this code here: https://github.com/xviniette/AsteroidsLearning

how can i save it's progress so i can resume to that state in the future?
what objects needs to be persisted in a database?
or is it only the weights?

>>57786291

Kek. Now that's a paper I would cite.

>>57762755
http://neuralnetworksanddeeplearning.com

Good site brah.

I'm setting up caffe right now.

>>57785674
Thanks for the response. I would like to use Theano to make use of my NVIDA 750M in an optimize manner. Where can I go to learn to set this up for the very first time? I want to be able to turn off cudnn as you said.

Since image recognition is solved what are some open deep learning problems then? I want to start narrowing down a focus before I apply to graduate schools and still have time to read papers and get my hands dirty

>>57786670
As stupid as I think chat-bots are: chat-bots. Facebook has been poking at this for a while.

Or anything involving properly using external memory. There's a couple promising starts but no winning solution yet.

Also playing platformers

File: 1479568016510.jpg (997KB, 2448x2448px) Image search: [Google]

997KB, 2448x2448px

Will I get better at machine learning if I put on cute socks?

>>57786979
of course the more layers the better

>>57778103
>dat image
It doooooo

>>57780728
>proprietary data
Uhhhh...

>>57786979
Only if you kys

How can SMOTE or ADASYN generate more than one synthetic neighbour if they work with categorical/nominal data? Do they just spam same instance k times?

>>57786457
If you want a good introduction to it, https://karpathy.github.io/2016/05/31/rl/ really helped me get the basics of it. He also provides working code at the end (mostly from scratch, only uses numpy) and links to explore further.

>>57786511
Generally you only have to save the weights (and shape if not specified in the code itself). No need to persist to a database, generally just dumping to a file works fine, especially when you're just getting started.

>>57788610
I was thinking of doing my 4th year CS project on something like this. Maybe using some kind of reinforcement learning method to learn a simple game like Pong or Tetris. Most of the research I've done so far has a lot of praise for DQN but this mentions that it's quite outdated these days, is that true?

>>57788764
I'm no expert in RL but I'd trust Karpathy in that respect. That's one of his main ffields at OpenAI now, I believe. You could go on ArXiv and search for the most recent RL papers too and see what they use.

Should Machine Learning / Deep Learning still be hot in 6-7 years? I want to get a PhD in it.

>>57790247
It should be. But by then, machine learning would be completely different than what it is now.

Just as 6 years ago, machine learning wasn't the same as it is today.

Completely different is a bit of an exaggeration, honestly. More like "some technique was discovered that everyone was excited about and thought this is the algorithm that would solve every question we have about general ai"

Just like convolutional neural networks are right now.

>>57790311
ok cool thanks