ANNs

I want to start learning about artificial neural networks, but my brain is pretty shitty and I can't understand anything (The only information that I have are various Youtube videos).

Do you anons know what should I study to learn more about them?

Attached: ann.png (294x172, 11K)

Other urls found in this thread:

cs231n.github.io/.
github.com/oxford-cs-deepnlp-2017
web.stanford.edu/class/cs224n/
twitter.com/SFWRedditImages

All you need to know is that you write data to input layer and you read data from output layer.
Nobody gives a fuck about what happens inside and you shouldn't either.

how similar is it to programing in C+?

Play with tensorboard

>how similar is it to programing in C+?

Attached: brainlet1.png (621x702, 56K)

Are you hoping the ANN will be a substitute for your "shitty brain"? How are you going to substitute an ANN for your shitty brain if your brain is too shitty to grasp ANNs?

How familiar are you with algebra, derivatives, and numerical methods?

You pretty much need to be a genius or have years of expertise in ML to coherently design new NN architectures.

Find a library for something like image processing and implement it in your language of choice. You won't program in it, you "train" it by giving it a sample dataset i.e. a folder of pictures of cats and things that aren't cats and label each in some way as cat or not a cat. You hand that folder to your NN, and it reads the image, checks to see that it decided cat vs not cat correctly, and repeats ad nauseum. Then you can theoretically hand it any image, and it will guess whether it has a cat in it or not. There's a lot of high level calculus that goes into the internals of adjusting the NN when it's wrong, and your ML library will handle it, you don't need to think about it. To you, a neural net should seem like magic, because it is. Just learn to implement someone else's library, and use it to do some simple image recognition. Then when you have a good grasp, try to come up with some other uses

not OP, but I find that most of the literature on ANNs forces me to go back to the algebra book, nobody ever has written a book on ANNs with both algebra and some code examples to go by in order to simply comprehension. Well, maybe Simon Haykin's book has some Matlab examples, but they're only available for educators.

uDemy's "Deep Learning: A to Z" is a pretty good broadspectrum course on the subject.

>artificial neural networks
As opposed to genuine neural networks?
I don't know what this means, but then again, my background is in computer architecture, not meme shit like data science.

You see, there are those new things called 'brains', you might have heard of them...

Attached: genuine_neuron.png (645x729, 77K)

>buying into the ML meme

I fucking hate you people. Fucking nerds treating all this shit as a black box is ruining the field. I wish this stuff wasn't as accessible with shitty 5 minute tutorials all over youtube so it couldn't be ruined by sub 120 iq dudebros and autists that think they're smart.

Phd candidate in ML here

If you're familiar with linear algebra start with something like this cs231n.github.io/.

Once you have a general idea on how NNs work, try implementing a MNIST classifier in a framework of your choice (I use tensorflow). If you get stuck at all look at how others have done it on gitlab/hub

Also don't listen to the black box advocates. Neural nets are perfectly understandable in theory, and there are ways to interpret a NN in practice (occlusion mapping, CAM, etc...)

Isn't it just logistical regression *n?

+99999999 /thread

Attached: mlmeme.png (542x729, 61K)

Yep, it's just optimising the parameters in a matrix so that it produces the result you want. It's not magic

I don't get why people are saying it's so difficult to understand.

What would be your advice for a student in bachelor's graduating soon but with shit grades cause he partied too much? I want to do a masters degree. Ive got work experience, so I was thinking of finding a job in RnD then using it as leverage, maybe even publishing a paper while I work. (My idea is to research the most effective ways change model predictions through data pollution in ways that is profitable for the adversary)

if your paper is mostly empirical rather than theoretical then you're part of the problem with ML

Honestly depends on your grades user. I don't think it will be too easy finding an interesting RnD job with a BSc, so best bet might be to go straight to a Masters if your grades allow (2:ii and up I guess?)

If not, go for the job, but do some nice side projects to show your potential supervisor.

Lol the problem will be all the hype around it. People are going to be so disappointed when they realise you can't throw ML at every problem

Im 2.0. I've already spoke with my supervisor about publishing a paper and he's interested. I can get an RnD job due to my work experience. 98% of analytics is building a system that can support it anyways

Honestly mate getting a first author paper published would more that make up for a shit GPA. Push for that, get the RnD job for a year, and apply to grad school for 2019 start

> Prove yourself

I'm an americuck going to college soon. If I want to get a job researching and developing AI, what should I do at college and what degree should I get? I planned on getting a bachelor's in CS and making AI projects on the side because there aren't many colleges with AI degrees but if any oldfags have a better idea, I'd gladly go with that. I just wanna make my AI waifu one day ;w;

Attached: 1501156830146.jpg (702x562, 46K)

> tfw physics degree
> tfw still went into AI
> tfw the degree doesnt matter at all

>Treating it as a black box
OP said he was new to NNs so yeah it's currently a black box to him, you want to start him off writing or understanding everything himself? Were you forced to write the Linux kernel or the C compiler before you used either? Did you have to design a processor, and be able to explain all of its componemts before you mashed your head into the keyboard writing that reply? No, you used them without thinking too hard about them, and learned more later if you were interested. Neural networks are complicated, pretending they aren't doesn't magically make them simple. It's better if op gets experience working with them, and then goes and examines how they work.

Physics had way way way more mathematics,statistics,probability and model real world experience over CS students.

Too many PhD physics become to machine learning,data analysis or Reseach CS.

Why not do a physics degree over CS then?

do CS and take stat courses on the side or something. Most of the fun things is in doing prep work on the data for easier recognition. the actual """"""""""""learning"""""""""""", does actually boil down to you feeding the prepped data, giving it to a bunch of different architectures with different parameters, and just taking the one with the best accuracy (of your choosing, if you're feeling devilish). Theoryfags are just salty that they had to take all the stat prereqs and in the end it all looped back into "we can tell you how these units work but in the end you just have to try out different shit and see which one works the best", which is pretty fucking funny.

CS is more easy and muh video games.
/sci/ laugh all day about CS students.

CS is easy? Maybe if you go to brainlet u in India

> compared to physics

I'm a ML researcher and I design new NN architectures (mainly RNNs for NLP)
The literal point of my research is to make black boxs that others can use
Fuck off stop ruining my field.

Now OP, start by watching videos on linear regression and logistic regression. These are much simpler. Once you fully understand them move to Neural Networks. You will find they are not much different.

Any papers user? ArXiv?

I have a paper I published on traditional ML, currently working on publishing some RNN stuff
Unfortunately I'm too scared to link myself here

New architecture? Do you have any tips on the publishing process? Clueless grad student here

you can program a NN in C++ you pleb

I also interested in this subject, mainly story generation, I want to process books and novels to create engaging stories, but all I find is algorithms that process cats and put nicolas cage face on trump's videos

Look into NLP. They can generate short news stories with the state of the art

github.com/oxford-cs-deepnlp-2017
web.stanford.edu/class/cs224n/

Once you understand neural nets you will understand what a fucking scam the current state of ML is. Even the advanced ones like RNNs and CNNs are fucking meme material. They have oversold these technologies so much that when the investors realize how """intelligent""" they are the field will enter another winter.

Obviously they are good for many classification problems and have revolutionized the algorithms in computer vision but 4 reel ANNs will not lead to AGI.

>AGI
Yes is science fiction today
>computer vision
Some problems in CV begin billons dollar industries.

Guarantee this will happen soon. I'm so shook being in ML

I feel like people aren't going deep enough. Why not have a network of massive nested RNN? We are like the component stage of electronics. Now we just need to figure out the computational logic and architectures required for huge leaps of advancements. You niggas are acting like what peeps thought of the computers in the 80s.

I greatly fear this tech