ok let's talk about programming AI.
Let's have a real thread on real shit, not your faggy fizzbuzz for your pajeet employeer.
I want to learn to make code that will take photo inputs and then make a new painting in some artist style.
What books or what resources do you have for that?
Install gentoo
>>55287056
Here, now fuck off https://github.com/jcjohnson/neural-style
>>55287056
lets start with a PhD in Computer Science.
Use this
https://news.ycombinator.com/item?id=11985709
Also, don't code your own lib if you actually want to "get shit done".
>AI
>real shit
Pick one and only one, faggot.
>>55287056
These links are ok
>>55287250
>>55287261
I'm more interested in evolving a general AI. Lack of computing power stops me.
>>55287285
>I'm more interested in evolving a general AI
Okay. I recommend getting a PhD under a professor with a similar goal. Then spend the rest of your life working with researchers on this problem
1a) Plan out a big data feed pipeline (aggregate painting images) to continuous feed machine learning, along with answers for learning the first batch
1b) Feed the data from existing analytics platform
2) Cloud architect out EMR/Hadoop/Clojure to etl data using scala, push it into s3 or another non nfs mount cloud storage
2b) Optionally push that into analytics platform like redshift to build regression to compare against performance of machine learning
3) Build multi-level machine learning algorithm in apache Spark (or Google TensorFlow) on cloud, load balance and distribute it accordingly for scaling such worker tier. Might need to look into color analyzer algorithms. Steal from Google DeepDream algorithm
4) Push output data to either s3 or to redshift/whatever and be ready to analyze the response, debug, repeat.
5) Once enough sample size for neural networking is performed in whatever cluster you have, push in real pictures and see performance in batches. Compare results with the data you have in redshift regression, or manually inspect them.
Above does not include messenger queue and cron scheduling you need to do.
If you didn't understand any of that above then I suggest you go learn your shit.
>>55287285
General AI is impossible, my humble opinion.
The only way to prove that it, infact, isn't impossible is to make one. Proving that it is impossible, is itself impossible, unfortunately.
Anyway, people don't even *know* what intelligence *is*, let alone making an aritifical one.
lol
>>55287264
/g/ hasn't heard about deep learning, I see.
>>55287285
yeah sure let me pull a tutorial for an evolving general AI from google
>>55287285
people do that shit on the scalable cloud nowdays retard
>>55287336
I have a simpler plan, one that has been tried albeit at small scale:
Create/find and integrate open-source versions of the following:
1) A physical simulator that is a compromise between performance and accuracy. A set of standard physical objects and an articulated creature with sensors and actuators capable of interaction with said physical simulation. Sensors and actuators are read/controlled via 2 float arrays.
2) A virtual machine with short, concise bytecode with the property that any bytestring is a correct program. A fast JIT (maybe LLVM) implementation of this VM. This VM should be able to host any arbitrary program that controls creature from 1) via actuators using data from sensors.
3) A modern genetic algorithm capable of running on multicores and clusters (probably natural evolution strategies or CoSyne).
4) A set of benchmarks set in 1) simulator that are designed to require intelligent behavior on behalf of creature controller software to solve them. Think about puzzles, navigation, labyrinths, simple games, pattern recognition. A set of benchmarks uses 1) simulator with creature controlled by arbitrary 2) VM software and outputs a fitness number, higher is better.
Then I'd like you to run 3) genetic algorithms on a sufficiently large cluster (perhaps sized in 100k+ cores and based on AWS, GCP or Azure) to breed 2) arbitrary VM software against 4) benchmark until it shows signs of general intelligence. Then email me, I'll inform you on our next steps.
You called yourself a god? Do your work then (^:
Also you could do all the above but also study deepmind's, openai's and FAIR's papers and bias your evolution to known good AI algorithms (this is what I'm doing for some long time already).
>>55287368
>Anyway, people don't even *know* what intelligence *is*, let alone making an aritifical one.
Laymen sure don't know what AI is.
On the other hand AI researchers have come up with a good, practical definition some time ago: "A system that can be trained to achieve complex goals in wide variety of environments"
This is a definition used by Deepmind and OpenAI as well.
The definition is here and you can run your Algorithm to score its intelligence.
Please don't derail this thread into empty philosophy.
>>55287458
Sounds like a research project you'd do with your research colleagues that I suggested.
Also, they will make fun of you for mentioning genetic algorithms. In academia we call it "stupid search"
>>55287406
Here are works on the subject by world-renowned research team: http://people.idsia.ch/~juergen/evolution.html
For example this one:
J. Koutnik, G. Cuccu, J. Schmidhuber, F. Gomez. Evolving Large-Scale Neural Networks for Vision-Based Reinforcement Learning
>>55287491
GAs are "stupid", a slow but global optimizer. Of course I'm not against applying other optimizers. There are dozens of metaheuristics and then some variants of gradient descent.
A variant of GA called CMA-ES was good enough to evolve videorelated, for example: https://www.youtube.com/watch?v=BJaVUTI008M
>Sounds like a research project you'd do with your research colleagues that I suggested.
I know this is a standard route to do research, but why do I have to waste additional 8 years on M.S. and Phd when I can do (and do!) all these experiments at home? My only problem is lack of computing power, but it can be solved by porting the simulation code to the browser and letting anons donate compute time by visiting my web page.
I have some code for such architecture but it isn't ready for primetime yet.
>>55287442
I can do this in the cloud if you will give me the money, anon. Cloud is not the cheapest source of compute though, crowdcompute is better. Hell, even a dozen of good old dual xeon servers in my room would greatly help here.
>>55287635
>GAs are "stupid", a slow but global optimizer
No, they aren't global at all
>>55287635
why this isn't used in games?
>>55287635
>I know this is a standard route to do research, but why do I have to waste additional 8 years on M.S. and Phd when I can do (and do!) all these experiments at home? My only problem is lack of computing power
Your other problems are lack of colleagues to work with and wasting time at a day job that has nothing to do with this research.
It isn't a waste of 8 years. You'll be doing this shit for 8 years, while publishing papers about it and making cool stuff.
And I seriously doubt your browser idea will work. Much better to use some school's or company's infrastructure
>>55287688
Some variants of GA are proven to be global optimizers, others are not. In practice GAs are known to find better optima than gradient descent.
Even deepmind has recently published a paper that used GA to achieve orders of magnitude compression of an autoencoder.
I'm not saying GAs are the best, but they may be useful for some hard problems.
>>55287741
>Your other problems are lack of colleagues to work with and wasting time at a day job that has nothing to do with this research.
I know some programmers/researchers, I discuss these and related things with them. I don't need any help in programming my stuff though. Also I don't have to work 100% of my time, right now I'm comfortably living off my savings lol.
>It isn't a waste of 8 years. You'll be doing this shit for 8 years, while publishing papers about it and making cool stuff.
I have seen and read about harshness of PhD life, I don't want to put myself into such a cage. Low pay, overwork. It isn't worth it. I understand that this undermines my "research" credibility, but who cares if it is just a hobby of mine?
>And I seriously doubt your browser idea will work. Much better to use some school's or company's infrastructure
It already works, though of course you need to run long simulations on the clients to recoup time lost on network IO. With asm.js and wasm one can reach close to native performance.
Anyway, I'm not saying I'm going to create strong AI, this is just a hobby of mine. If you are interested we can discuss it via email: [email protected]
>>55287716
Because games spend almost all compute time on graphics and don't really have time to spend on simulation and AI. Market cares about the graphics most.
>>55287483
OP asked about, and I quote, general AI.
>>55287904
>right now I'm comfortably living off my savings lol.
>My only problem is lack of computing power
Why complain about issues that are solved with money?
>>55287994
>Then why are you complaining about issues that are solved with money?
My savings won't buy me much power, and it is stupid to waste your last money on a hobby.
>>55288009
I answered with definition of general AI commonly accepted by researchers.
Denying it and talking about "consciousness" and "true AI" are plebe-tier speculations. Please leave philosophy at /lit/
>>55287056
Ok let's talk about programming AI.
The things to look into are machine learning and whatever things like genetic algorithms come under.
They teach this in undergrad CompSci.
Get the main books.
After you understand the basic algorithms you'll be able to work out how to make it.
You can look into libraries like Weka and ScikitLearn.
Then go to google scholar to check for enhancements you can do.
It's not exactly hard but it might take you a year if you're starting from scratch with AI.
>>55288408
This is a not bad advice, I approve. Also read r/MachineLearning.
Is there any reason for me to not use Caffe on Spark for general DNN stuff? I have access to a ~25 node cluster
>>55288451
Caffe is widely used in industry.
>>55288633
For some reason nobody uses it in the area of research I do, even though AI/ML practices are widely used. Probably because most people are still using nfs's instead of dfs's.
>>55288701
Caffe isn't used in research because it is rigid, with it's JSON config language. Tensorflow is much more dynamic and meshes well with scipy/numpy.
bumpu
Why don't we start an IRC chat on AI?
>>55287056
How does that have anything to do with AI?
If you want to use a machine learning approach just for the novelty, then you could try to implement that one program that tries to map triangles over an image and judges each triangle's fitness
>>55289580
Deep Style Transfer is a cutting edge deep learning application, so it has its place here.
>tfw no AI gf
