Imagine developing INTELLIGENT programs... isn't it better than making jQuery resizable boxes LOL ?

It always makes me chuckle that AI programmers are trying to program themselves out of a job.
Once it is nailed down, they'll be replaced by what they designed.
Personally, I think AI in a broad scope is a pipe dream, it's an overrated reality that people have turned into a buzzword.

>>53470742
This is true. The last programming language will be English.

In the not too distant future, we'll be able to program using English.

This is coming soon.

Sure, some will still use ASM, C, etc. buit that will be just for low level stuff

Lot of stuff will be done in english

>>53470797
Possibly. We've already seen plenty of advances at interpreter level.
I don't really see how that is a massive stride though.

>>53470687
The public release of tensorflow is dogshit and nobody is going to use it for the next 5+ years at this rate.

If you actually want to do ML, use torch or theano. Theano is more low-level, but it's easier to create new nets with it. Torch is more set-pieces-like and requires more work to make unique new models, but it's easier to get started with.

Also, >>53471982

>>53470797
bullshit, english is way too vague to properly communicate instructions to a machine.

it may have a few english semantics, but you'll never be able to just type what you're thinking into an IDE and have it spit back a program.

>>53470687
thanks senpai

>>53472188
It's one of the current "hot" areas of research. Not English-to-program but English-to-directions. There are several areas of focus. In robotics, it's action sequencing, like reading a recipe to a robot and having the robot follow the recipe for you. In nlp, it's typically the likes of question-answering, which can be used as a part of a English-to-program system.

I have a Fermii.

What GPU to get for TF? GTX 970Ti? Should I wait for Pascal?

>>53472263
You need sm52/cc5.0 or better as I recall for pretty much anything except theano, but theano is significantly faster with sm52/cc5.0 anyway because of cudnn.

>>53470687
>take quick course on AI
Kek, current AI(s) is just bullshit
You cant take a course on something that doesnt really exist

>>53472263
Also, to get reasonable performance or work on relevant models, you will need a titan or titan X. And don't bother with "science" gpus, double-precision float performance doesn't even begin to matter. It matters so little that the hot thing is using float16 instead of float32 now.

>>53472304
Luddites belong on >>>/reddit/

>>53472308
What's the difference between titan and geforce lines?

I'd want a card to play with TF, Caffe, but also some integer stuff (e.g. DES), for which last time I've checked AMD cards are much better, and from time to time I play some games.

>>53472357
Titans are Geforce GTX Titan. They're the same line. It's just the name of the flagships instead of just a number.

ML is never done with integers, except in special research toward that, where typically one uses fpgas to operate on integer or even binary weight matrices.

There is no ML framework that targets opencl, but there are branches of some frameworks, like torch (and I think caffe too), which do. They tend to be significantly slower than the cuda variant, however. Nobody really does ML with AMD cards so for all we know, even with the slowdown from the poor opencl implementation, using AMD cards, it could be faster than cuda + nvidia. I wouldn't risk it if I were you, unless you want to actually write opencl code to fix the performance issues.

>>53472452
I prefer nvidia, I'm only worrying about the integer performance for other tasks. I guess I could get 2 cards.

>>53470687
What path should I take to build an open source AI waifu?
Or at least something on par with Siri/cortana

>>53470742
This

>>53470742
Researchers != programmers.

>>53472529
Everything not made by company in ML is either open-source or open-spec.

>>53472588
A good programmer researches their own field.
This kind of opinion irritates me no end. You are not a ground breaking programmer if you only know what books tell you. That's not breaking ground, it is following guides.
However, I do agree to the extent that programmers can work from hypothesis that are devised from primary researchers.

>>53472624
Not sure if retarded or just pretending.

>>53470687
It's boring shit. The theory is all about training some abstract weights. Nothing visually interesting going on. I'd much rather work on graphics or games or something.

>>53472693
Well you're very stupid.
You obviously don't realize the (visual) possibilities that can be achieved with neural networks.

>>53472693
http://yann.lecun.com/exdb/lenet/translation.html
Shh baby shh

>>53470716
because I'm too dumb to do it. I'll probably be a rajeesh-tier coding monkey for the rest of my life.

Because ML can't into fizz buzz

>>53470687
I agree with you, op. But ML is at a level above ordinary programming. /g/ autists just can't understand this. Also they can't into math/numerics.

Say, 2% of people become programmers, and of these 2% are able to learn ML. We are a minority inside a minority.

>>53472258
http://arxiv.org/abs/1510.07211

>>53472529
1) find/compile a dataset of waifu behavior
2) train https://github.com/Kaixhin/Atari on that dataset

>>53472263
GPU is not a hard requirement, it just accelerates training 10-100. You don't even need any special languages if your experiments are small.

For example, I wrote feed forward neural net in JavaScript, trained it on mnist and it worked OK.

Deep learning can be simple.

The problem with ML career-wise is that without creditionals (M.S. or PhD) you won't be hired. So, its just a hobby of mine.

>>53472693
>I'd rather work on games
Fucking lol

>>53473445
I know there are people who actually research this, but it's not as hot as "language understanding".

>>53473554
Yeah, but you can get 0.6% error rate with a "dumb" CNN in less than a day only with a GPU. You'll be stuck at what, 2% or so in FC by training twice as long? (Haven't actually touched FCnets on mnist in years, I don't recall). Convnets in software is just a lost cause.

>>53474144
The best I got on my feed forward network (simple mlp, not even a CNN) was 1.8-2%, yup. In principle for some tasks it is enough, but I'll improve it anyway.


GPUs are cool for CNN's (100x), for rnns its more like (10x). If you really need it you can use aws cloud, there are GPU instances and tensorflow images for them.

Most of the work, at least initially, is getting a good dataset, which depending on what you want to do might be virtually imposible. It's just boring and tedious work. And if you are using a public dataset it's full of mistakes.

>>53474746
Most public datasets are gr8, nig. But if you need a custom dset, yeah, you'll typically end up using 90% of your time handling the dset and 10% of the time actually doing research/development.