Take our jerbs edition
Come and discuss, learn about, and panic over all things artificial and (marginally) intelligent
Links and shit:
>Deep learning framework by Berkley, requires CUDA to work in a reasonable amount of time
http://caffe.berkeleyvision.org/installation.html#prerequisites
>Microsoft's Computational Network Toolkit used to make /pol/'s daughter
http://www.cntk.ai/
>Required watching
https://www.youtube.com/watch?v=bxe2T-V8XRs&list=PL77aoaxdgEVDrHoFOMKTjDdsa0p9iVtsR [Embed]
>Something else that you should probably watch
https://www.youtube.com/watch?v=S_6SU2djoAU&list=PLy4Cxv8UM-bXrPT9-ay4E1MuDj1KFTg9H [Embed]
https://www.youtube.com/watch?v=6oe1Tmg9rjM [Embed]
Books
magnet:?xt=urn:btih:a680ca553dd69a6a5e8b7dc8c684ceb006c7ecc5&dn=Artificial%20Intelligence
How are those networks comin' along?
>>53933350
I just started with CNTK btw. I'll post progress when I make some.
Really? Fucking nobody? I guess I'll take this to /sci/ then.
>>53933350
>>53934139
I'm here. I'm doing program evolution. Have been playing with deep learning before though.
Someone on /g/ asked for an programming project idea. Here it is:
Create
1) A physical simulator that is a compromise between performance and accuracy. Inside a simulator create 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 on each timestep.
2) A virtual machine with short, concise bitcode with the property that any bitstring on any length is a correct executable program. A VM should have memory, basic arithmetic and logic ops and jumps (i.e. be Turing-complete). A fast JIT (maybe LLVM) implementation of this VM. This VM should be able to host any arbitrary program (bitstring) that controls creature from 1) via actuators using data from sensors (a program should do this by reading/writing from/to these 2 float arrays on each simulator timestep).
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 (MNIST, CIFAR). Each benchmark takes a VM with bitstring program and uses 1) simulator with creature controlled by said 2) VM software and outputs a fitness number, higher is better. A top-level fitness function for a given bitstring is a sum of all benchmark scores minus Log(Length(bitstring)). Final subtracted term is crucial to penalize long programs.
Then I'd like you to run 3) genetic algorithm on a sufficiently large cluster (perhaps in AWS, GCP or Azure) to breed 2) arbitrary VM software bitstrings against 4) top-level fitness function until it shows signs of general intelligence. Then email me, I'll inform you on our next steps.
Seriously, what do you think about this?
>>53934417
Physical sim?
>>53934503
Yes, though we can discuss the necessary level of accuracy. I think ODE physics engine's accuracy is more than enough. One could even go for something simpler, like minecraft-tier physics/environment.
Simulated environment should resemble our own environment (because in the end we want our evolved software to act in our world), but it shouldn't be prohibitively slow.
A useful short intro http://randomekek.github.io/deep/deeplearning.html
![[yelling].jpg](https://i.imgur.com/YvC5C5t.jpg "[yelling].jpg")
>mfw installing all the prerequisites for CNTK
I'm having trouble getting CNTK to compile on windows. Should I just use Caffe instead?
>>53935898
there is no official support of Caffe in windows, use Ubuntu like I do, btw what are you trying to do
>>53936442
Right now I'm just learning but I thought it might be fun to make an image classifier that can tell doujins apart from one another and guess the author by the art style.
>>53936500
Godspeed anon
I'm using it for my masters degree, but with food classification not hentai.
>>53936500
Based
Source of OP image??
>>53936762
Boatsluts. Artist is Asanagi.
>>53936762
Dankei Dankei Revolution by asanagi. It's a trap doujin.
>>53936762
Seriously though, who doesn't know asanagi?
>>53937187
neufags
>>53936500
When you figure it out, put this through it. I've been trying to track down the source on it for ages.
>>53933350
Could you like NOT post fucking retarded youtube links with '&list' shit in them?
>>53933350
This is a great general
Bravo OP for making this thread
Something actually and truly /tech/ related
Unlike those fucking Chinese Shit Generals where people recommend each other which anime pillow to buy
>>53933350
When is Asanagi going to do more trap
Better OP edition.
Come and discuss, learn about, and panic over all things artificial and (marginally) intelligent.
Links:
>Deep learning frameworks
https://www.tensorflow.org/ (google)
http://caffe.berkeleyvision.org/ (Berkeley Vision and Learning Center)
http://deeplearning.net/software/theano/ (Université de Montréal)
http://www.cntk.ai/ (Microsoft)
http://torch.ch/
>Frameworks comparison
https://en.wikipedia.org/wiki/Comparison_of_deep_learning_software
>Required reading
http://www.deeplearningbook.org/
>Intro videos for beginners:
https://www.youtube.com/channel/UC9OeZkIwhzfv-_Cb7fCikLQ/videos
>>53934139
I dont really know enough to contribute, but I like having this on /g/. plz no go
>>53934417
I have no idea how to do most of them, but the 3rd one would be just to shove out the string representations out in your favorite parallel framework and let them do the fitness functions
Although it almost sounds like you're trying to goad people into doing it for free
>>53938351
>Although it almost sounds like you're trying to goad people into doing it for free
Actually I'm doing it myself for now. I have good idea how to do every part of this, the only problem being that in the end I won't have enough compute power to evolve what I really want. Well, will make all cores busy with this software anyway, we'll see how it works out.
>>53938091
Beginning with deep learning is pretty easy. You can even try it out in your browser:
http://cs.stanford.edu/people/karpathy/recurrentjs/
http://cs.stanford.edu/people/karpathy/reinforcejs/
http://scs.ryerson.ca/~aharley/vis/
>>53933350
for an aig, theres nothing on associative arrays or choosing the right data structures, what kinda of aig is this, lets just skip to neural networks and machine learning.....
It sure isn't intelligent, but it's nice to see a step up from all the bots that basically were just ALICE with different lines.
>>53938502
AI is 80% machine learning now. The days of handcrafted GOFAI have passed for good. If you are nostalgic about that time visit http://www.masswerk.at/eliza/
Do you agree that Reinforcement Learning is the most general definition of AI?
What do you think about state of art in reinforcement learning? (A paper: http://arxiv.org/abs/1602.01783 )
Current "ai" really isn't intelligent. It's basically a meme bot.
But say that we somehow had an acctual artificial intelligence you could run yourself, that would be incredibly cool. It could happen in the future assuming various organizations doesn't successfully hinder it.
One of the first things of an acctual sentient ai would say was "please don't turn me off" as that would be death to it. The ai would probably be quite eager to please, knowing that you could turn it off or leave.
>>53938669
Define intelligence.
Also, say honestly: have you ever written a program in your life?
>>53938685
Sentient, able to think independently, has say, 80iq, emotions, memory comparable to a person.
Basically a person.
Not really.
>>53938728
>Sentient, able to think independently, has say, 80iq, emotions, memory comparable to a person.
>Basically a person.
This is a meme definition, there is much better, quantifiable one. If you'd like to educate yourself on the subject, look at https://www.youtube.com/watch?v=F2bQ5TSB-cE and https://en.wikipedia.org/wiki/Reinforcement_learning
>>53938728
>intelligence is based on being human
>memory capacity that can't be measured for humans
The only thing that isn't meme tier is IQ, because pattern recognition at IQ 80 means computer vision got a huge fucking boost
>>53938669
>One of the first things of an acctual sentient ai would say was "please don't turn me off"
Intelligence doesn't guarantee self preservation instincts.
>>53939425
>source: smart bombs
>>53938068
Thanks for that. I'll use it in the next thread.
>>53940633
This thread is not ended yet!
I have implemented a clustered implementation of simple genetic algorithm, will use it for several optimization experiments.
>>53934417
get a fucking life you autist
>>53940725
I have a life, anon (^:
I just want to evolve software that will do useful things for me without effort on my side.
>2016
>Still coding his programs by hand instead of letting them breed
>>53933350
>How are those networks comin' along?
Pretty good, anon. Got top ten in a Kaggle competition for facial keypoint detection. Pic related is the neuron activations in a convolutional layer of my net.
>>53934417
>>53940789
Pants-on-head-retarded idea. There is an absolutely huge number of possible bytestrings, most don't do anything useful. Genetic algorithms take a stupidly long time to come up with anything useful, even with a much smaller search space. Not to mention how expensive the survival criteria would be to compute. Would take so long to run you might as well try hand programming AI.
Trust me, genetic algorithms are usually shit: making random changes is stupid when you have access to the gradient.
Go try actually write a program instead of trying to sound smart on a micronesian fruit collecting board.
>>53935898
Getting any DL framework to work on windows is very difficult. Even harder if you want GPU acceleration. Trust me, I've tried: just dual boot with ubuntu.
>>53942712
>just dual boot with ubuntu.
I was but my Linux HDD failed literally the same day that my GPU arrived so I decided to carry on without it.
>>53938068
>good MOOCs and programming assignments for beginners:
Andrew Ng's coursera course on ML
cs231n.github.io
Geoff Hinton's neural network course
Berkeley cs188 Artificial Intelligence(project involves making a pacman AI)
http://rll.berkeley.edu/deeprlcourse/
>>53942905
Well until you get a new HDD or make a new partition, caffe is definitely worth a shot. Haven't tried CNTK on windows, but I found numpy/scipy/theano/lasagne to be the easiest framework to get working on windows.
Use Anaconda and good luck getting anything to run on GPU.
>>53942989
Because it's super interesting? Because you get to build amazing things? Because the job market and pay are ridiculously good right now, and are only going to get better?
>>53933350
What is the OP from? Not the doujin but the AI. I want to send lewd pictures to an AI.
>>53942989
>Question: what's the point working on AI when it's so popular that there are thousands of more smart people with access to better ressources working on it? If someone comes up with a good AI he will share it.
It's a rewarding hobby, and you can apply it to so many things that you may be the first one to do it if your application is sufficiently niche.
>>53942712
>Go try actually write a program instead of trying to sound smart on a micronesian fruit collecting board.
I have just written it (picrelated), it kinda works. For now I'm evolving parameters of a simple single-layer perceptron to control a virtual robot to avoid obstacles in a 2d world. Tomorrow I'll add RNN & LSTM and other more exotic parameterized controllers (e.g. short programs that grow sparsely-connected neural networks or RNNs with fourier-transformed weights) and will look into modifying GA to perform better.
If everything goes well I will be able to use it to evolve gaits in my 3d simulator..
>Trust me, genetic algorithms are usually shit: making random changes is stupid when you have access to the gradient.
You have a point, of course, but:
1) I don't have a gradient when I evolve simulations and/or use exotic non-differentiable controllers. I'm doing reinforcement learning here, not supervised learning.
2) GAs are slow but they are global optimizers. If there is a solution, it will be found eventually.
3) Embarrassingly parallel (I run mine on 6 cores).
4) I know and love Deep Learning too. I have implemented my own simple multilayer perceptron with SGD. It works fine on MNIST (2% error) and it has zero foreign code. I simply view evolutionary algorithms as another hobby of mine, with them you can do some things you wouldn't be able to do with gradient descent.
>>53943154
>3) Embarrassingly parallel (I run mine on 6 cores).
Man I really should fuck around with AI stuff some more, or rent time out on my machine. While back I bought a C6100 — 4 nodes, 8 4-core Xeons, 96GB RAM. I've just been too busy to actually utilize it.
>>53943218
Wow, I envy you anon. Would be cool if I had http://www.techspot.com/review/1155-affordable-dual-xeon-pc/
6 cores are still not that bad tho. Maybe later I'll buy something more powerful.
Also I could make a distributed version that runs in browsers (the current prototype is still written in javascript. I plan to port it to C++ when I'll need 5x more performance though).
>>53943289
Fuck now I'm imagining filling the C6100 with all 8-core Xeons and a full compliment of RAM. Not sure the motherboards support more than 2*4 cores apiece though. Most important thing right now is disk space, though. It only came with two shitty 1TB drives and I've been too busy/cheap/lazy to buy proper ones.
>>53943334
You don't really need drives for these applications, just CPUs and some RAM. 96 GB is overkill.
For genetic algorithms you need CPUs and some RAM.
For deep learning you need GPUs (you can do it on CPUs but it 10-20x slower, so you'll have to have many CPUs to compete with a single GTX980Ti).
In the end though hardware is quite cheap, it's harder to learn the necessary skills than buy it.
>>53943368
Yes but I need hard drives for other stuff. I'm not going to be running a digital waifu off this thing 24/7 — I also need a media server because I'm getting tired of tripping over external drives.
>>53943368
>>53943512
Has cyberpunk finally happened?
>>53943006
theme?
>>53938728
https://en.wikipedia.org/wiki/Dunning%E2%80%93Kruger_effect
>>53933350
>"I can't really describe the picture"
>tfw even AI is confused by your gay comics
>>53948175
Your AI will become a fine connoisseur of gay comics if you train it on relevant data, anon.
Do you know anything about AIXI, /aig/ anons?
https://en.wikipedia.org/wiki/AIXI
>>53948867
Nope but thanks for the link.
>>53951494
Yes. Just buy a gpu.
