Hey /g/ anyone work with artificial intelligence in any way shape or form? I find it fascinating and think that's the field I'll aim towards in life.
>>53561480
AI is kind of a dead field right now. Why not look into neuroscience instead?
>>53561480
just got done watching terminator did you?
AI is fairly boring and quite a stagnant field with most of the development being focused on optimising techniques for ever more niche applications.
>>53561480
The traditional type of AI is gone.
It's all machine learning/deep learning these days.
brush up on your math, especially multivariable calculus, linear algebra, and probability
Well, I think you should pursue your dreams.
Just remember that it will take a lot of hard work and you might not end up where you originally aimed!
- A happy anon who wants happiness for all anons
>>53561480
I've done some work with neural nets. I found it to be very interesting, and you can do some very powerful things with them, but they don't scale very well.
If you are looking to get into AI, you should manage your expectations. It is a lot more theory and advanced math than making your own Jarvis.
>>53562071
>A happy anon
You've come to the wrong site motherfucker
>>53561981
>AI is kind of a dead field right now.
AHHHAHAHHAA
>>53561480
I have a MS in intelligent systems (AI+ML). I do NLP ML work for a living.
I'm actually moving away from the field. The results are interesting, but day-to-day there's a lot of data management since all the code has already been written and it's just a matter of configuring the process the right way and run all of the data through it.
>>53562071
>Well, I think you should pursue your dreams.
>Just remember that it will take a lot of hard work and you might not end up where you originally aimed!
>- A happy anon who wants happiness for all anons
>>53561480
I do deep learning as a hobby. I have my own simple library and I use torch and theano sometimes.
Also I play with reinforcement learning and simple genetic algorithms. That's a pretty rewarding hobby, I'd say.
To an uneducated person Machine learning looks like magic, but you can learn how to use it if you are motivated.
>>53561981
Machine learning is literally the hottest field right now with billions of private & state investment, anon. I hope you are just trolling.
>>53562034
>stagnant
>4 days ago deep-learning based AI beat world champion in ancient game of Go
>>53562287
Sounds more like 'classification systems' to me. Do you have any proof of any *real* AI systems?
>>53561480
depends on how far you want to go
you can build stuff using frameworks and libraries available, then it's just plugging in the data you gathered
or you can truly want to understand the field, then you need to know your math inside out
>>53562323
Reinforcement learning is real AI. Given strong enough reinforcement learner you can reward/punish it into solving any problem a human can solve, and possibly beyond.
You'd rather listen to what prof. Hitter has to say about the subject, not me: http://m.youtube.com/watch?v=F2bQ5TSB-cE
>>53562316
>finally got good at something humans were great at thousands of years ago
Only a few more millennia of innovation to go!
>>53562323
What do you mean by "real AI"? Categorization is a classic AI application.
If you are talking about artificial human brain AI, that doesn't exist.
>>53562287
How good do you have to be at math to get the basic concepts?
>>53562323
>*real* AI systems
Define a real AI system. Without mentioning "consciousness" or similar BS.
>>53561480
If you are motivated you can download this book http://www.cin.ufpe.br/~tfl2/artificial-intelligence-modern-approach.9780131038059.25368.pdf and skim over it (because reading it whole is too hard, even in universities they don't study the whole book in one course).
Learn some programming, any language will suffice, but many people use python or lua (C may also be useful. Knowing programming is really one of the least significant problems when learning about AI.
AI programs can be written in any language, for example I have implemented my computer vision codes in C++ and my neural network library in javascript (just because JS is reasonably fast and browser environment is convenient for debug/io/plotting). Btw here are bitmaps of weights of one of my simplish neural networks trained on mnist dataset. This is almost "hello world" of deep learning.
That all being said, knowing some ML on amateur level isn't that hard, especially if you are a strong programmer like me. Being a professional ML practitioner is way harder, because you have to know much more theoretical facts about algorithms you are using, to have real confidence in your results.
>>53562482
I'd settle for an example of AI that can emulate a fly brain.
>>53562574
We aren't there yet.
>>53562482
By real AI I mean General Artificial Intelligence - a system (learning agent) that can be taught to solve any problem a human can solve.
There is a wide consensus that Reinfocement Learning is the definition of this problem, and that strong reinforcement learning agents can solve very hard (even human-level-hard) tasks. By strength I mean numeric measure of agent's performance on test suite.
For example, current state of art in Reinforcement Learning is Deepmind's A3C agent ( http://arxiv.org/abs/1602.01783 )
Previous state of art in RL were variations of deepmind's DQN agent (the one that famously mastered 40 atari games). This one has a lot of implementations, for example https://github.com/Kaixhin/Atari
>>53562643
>General Artificial Intelligence
Read SciFi if you want that.
>>53562643
Those examples are really advanced, but they are nowhere near General AI
>>53562574
You have to define precisely what "emulating" a fruit fly brain is.
For example it is known that in ~2011 there were devised first superhuman visual pattern recognition systems: http://people.idsia.ch/~juergen/superhumanpatternrecognition.html
Machine Learning systems already outperform humans in some domains, without emulating the intricacies of the human brain.
>>53562664
That's a pretty mainstream term nowadays:
>Speaking at a recent Google ZeitgeistMinds event, Hassabis said that DeepMind had set itself up as the Apollo Programme for AI - referencing NASA's efforts to put a man on the moon - with its primary mission being to "solve intelligence and use it to solve everything else" through general-purpose learning machines.
http://www.ibtimes.co.uk/demis-hassabis-deepmind-apollo-programme-artificial-intelligence-1501052
>>53562574
I'd say winning at go, trading stocks and reading and comprehending natural languages to some degree are better AIs than fly brain emulation.
As for the brain emulation, if I'm not mistaken, scientists are basically waiting for quantum computing so they can literally copy brain structure and get AI without really knowing how it works.
>>53562721
So far 3 x 5 = 15 is the goal to beat.
>>53561480
After the ai "winter" ai was renamed "machine learning" and is a field of a few different disciplines such as NLP and computer vision. If you're interested in AI, I would suggest picking a specialty
>>53562680
What I'm saying is that Reinforcement Learning is a precise definition of "General AI" problem.
Of course particular algorithms designed for solving this problem (table Q-learning, DQN, A3C) vary in performance and generality. But still, there is a quantifiable progress (see the graphs in the papers) in both directions.
DQN is already general in a sense that a single algorithm can solve 40 different atari games without modifications, using just pixels from screen & rewards as inputs and outputting virtual button press events as outputs.
A3C is even more cool, the same algorithm is able to learn to
1) Navigate a maze to fulfill a complex objective
2) Drive a virtual car
3) Control a virtual robot in various motor tasks
It clearly looks like some degree of general learning capability.
>>53561981
In actuality the field of ai has somewhat stagnated and is looking toward biology for help. So... this.
>>53562785
You are wrong. Last 5 years have brought us multiple breakthroughs in machine learning and billions of investments. Biology has little to do with it, modern deep learning systems take no more than a rough inspiration from neuroscience.
If having algorithms that may summarize content of arbitrary pictures in natural language and even answer questions asked about the contents of the picture is "stagnation", then lets hope there will be even more of it.
>>53562890
Hahaha
>implying we're not just copying things from the 70s.
>>53562943
Neural nets are old. Only hardware has advanced.
>>53562890
>what is static image pattern matching?
>>53562943>>53562966
You guys clearly don't read the news/papers.
The wave of architectural innovations in NNs is mindblowing. I can't even say I know all the major architectures anymore.
for example this one https://youtu.be/hVv4M0bTBJc?t=430 and http://arxiv.org/abs/1603.01417 and a ton over weird architectures. Some people call it "a cambrian explosion in deep learning".
>>53563000
>clearly haven't even tried to write an algorithm that captions images.
do you even program, /g/ ?
>>53563085
impressive pattern matching anon, but far from AI (in the general sense).
http://arxiv.org/abs/1510.07211
>On End-to-End Program Generation from User Intention by Deep Neural Networks
>This paper envisions an end-to-end program generation scenario using recurrent neural networks (RNNs): Users can express their intention in natural language; an RNN then automatically generates corresponding code in a characterby-by-character fashion. We demonstrate its feasibility through a case study and empirical analysis. To fully make such technique useful in practice, we also point out several cross-disciplinary challenges, including modeling user intention, providing datasets, improving model architectures, etc. Although much long-term research shall be addressed in this new field, we believe end-to-end program generation would become a reality in future decades, and we are looking forward to its practice.
codemonkeys soon to become redundant (^:
>>53563116
This is clearly way beyond pattern matching - to answer a question about pixels the networks learns a hierarchy of features and an algorithm (running in embedded RNN) to produce answer by focusing attention on various parts of image.
Pattern matching would be if we just took all possible objects, matched them with every possible position on the image by sliding window + mean square error, then took the list of obejct positions that passed the threshold, then run it through some handcoded program to produce answer from that list and question.
That would be pattern matching.
This program was trained on (image,question,answer) triples and nothing else, and it is shown that it generalizes to images and questions it has never seen.
>>53563211
bullshit. Humans tagged the images that were used for training. Pattern matching. Nothing else. Still impressive. Not intelligent.
>>53562507
>How good do you have to be at math to get the basic concepts?
Basic calculus is enough. Knowing what a matrix and a vector is and how to multiply them is necessary. Understanding how a gradient descent works is necessary.
A general understanding what is supervised, unsupervised, and reinforcement learning.
But all these things are pretty easy, if you are motivated you'll pick them in a month at most.
To use standard ML frameworks/models you don't even have to know all that, you can treat your model as a blackbox and stuff your data into it. It may work, it may not work.
>>53562507
Calculus, and linear algebra.
>>53563313
Well, humans learned from their "tags" that were given by their parents/relatives as well.
Definition of "pattern matching" can be pretty arbitrary. I prefer to draw a line by the level of complexity/nonlinearity of computation used to produce an answer: thus the simple sliding-window MSR approach is pattern matching, but a nonlinear multilayered/multistep computation that is learned by NN is not.
You may have another definition though.
>>53563313
https://en.m.wikipedia.org/wiki/AI_effect
>2015. Darpa robots still remote controlled.
https://www.youtube.com/watch?v=8P9geWwi9e0
Humans still decide what the machine is going to try. The things it is going to try is based on pattern matching. This is part of what is going to be necessary for AI, but we are still hundreds of years off from even semi-intelligent machines.
>>53563388
Robots and AI/ML are pretty separate fields, with various approaches to controlling agents. Only recently there has been a push to apply learning to robots. Robots are hard for ML because they are realtime, they violate IID assumptions, and they are highly nonlinear dynamical systems, also the iteration/experimentation rate is slow.
Still, I have a small robot I have built. Not really anything AI in it, just inverse kinematics for slow walking and opencv for ball following.
OP, I would never tell anyone what they should do with their life, but if I could do things over, here is what I would do:
I would go the extra mile with EE and Neuroscience courses and try to find a project like Blue Brain to contribute to. These hard problems will never be solved with purely analytical methods. The future will be determined by numerical methods developed around physical measurements of brains. AI from a software only perspective is dead.
>>53563555
>ai from a software perspective only is dead
This.
Do neuroscience. Do yourself a favor. Computer Science is dead. Applied math is dead. Norvig is dead.
>>53563632
>>53563555
But its not dead
>>53562890
>>53562410
>>53562316
I like my neuroscience, I have actually read several dozens of papers about it. But it looks like deep learning is the future, it's simple and its limits aren't known yet.
First airplanes were not developed by ornitologists. First cars were not developed by horse scientists, etc.
>>53563632
AI from a software-only perspective anon... big difference. The future of AI will be determined from building a brain from a numerical methods project. It will require many disciplines to contribute over many years, but it is the only way to really advance intelligent machines to the next levels.
>>53562527
That's F'n Awesome! Some real insight, there! Thx.
>>53563700
>>53563555
I agree that numerics is very important. Deepmind, FAIR and MSR have large engineering teams that make sftware/hardware inftrastructure to run big computation graphs.
This isn't directly related, but I believe this guy is capable of bringing all the various sciences together to create some new models that will take things to the next level. No, I have nothing to do with this project, but I do believe these guys are on the right path.
https://www.youtube.com/watch?v=3M4UgeLW1cI
>>53563762
Human brain emulation is a real possibilty. There are already 10nm res scans of 1mmx1mmx1mm of mouse cortex: http://www.cell.com/cell/pdfExtended/S0092-8674(15)00824-7
The process could in principle be scale to whole mouse brain before 2025. Then cat brain, then ape brain..
But this project will be large and slow. With current success of deep learning we may have subhuman-but-general RL agent in a couple of decades, so before whole brain emulation.
That's just speculation of course, read and think for yourselves.
>>53563762
Interesting video, anon.
I was planning to go for Neuroscience through Neuropsychogic behaviour and learning speciality, should i switch over to CS, math, engeneering and the like?
I could switch to nanotechnology for free as well, and i'm pretty much aiming to work on r&d on biologic-machine interface but not so sure of the actual positions for work.
>>53563942
No! Do Cs and neuroscience if you can
>>53563942
>biologic-machine interface but not so sure of the actual positions for work.
Yeah, it won't be a viable field for jobs for quite some time. Right now, corps like google who have enough money to waste on this will be able to play around with AI, but other than that you won't have many other opportunities.
>>53563906
Yes, with this type of approach, they will be able to easily test new drugs, etc to develop new analytical models that work as expected. Anything short of this approach is just throwing darts in my opinion. This type of research is worth spending some money on.
>>53564197
On the other hand I see promise in achieving even subhuman general AI, training it in applied sciences, then running thousands of instances in the cloud to solve humanity's problems. Looks like deepmind has this as their endgame plan.
These AIs don't need to be smarter than human, they can outcompete human scientists just by sheer numbers and rapid accessto databases.
Just several more improvements over A3C and this may become real.
If we won't use AI to do science the progress will be much longer, humans take 25 years of training just to start doing some independent scientific work.
>>53564197
The issue with simulations of the brain is that it would be computationally expensive to perform drug screening. I'm currently running some simulations using autodock4 on a supercomputer and it takes ~3 days to do ~1780 ligands to one receptor (and 3 different structures). Imagine the time it would take to simulate all pathways within the brain, while keeping true to the biology of the brain. These type of advancements are cool, but we need more powerful computers to make this work practically.
>>53564289
>I'm currently running some simulations using autodock4 on a supercomputer and it takes ~3 days to do ~1780 ligands to one receptor (and 3 different structures).
Very cool. Did you think about using this data as a ground truth for training predictive models? You could speed it up by 10-100x. Everybody in big pharma does so.
>>53564289
Understood. This is why I think true AI is a still a long way off. We still don't understand the brain and current supercomputers aren't enough. Maybe if Moore's law holds up for a few more years? Regardless, it is still the right approach in my opinion.
>>53564387
Holy shit, pay attention to current research. This is what the government is currently investing in.
>>53564339
No idea about industry, I'm just a student turned employee at my university (although I hopefully get started with my PhD soon).
AutoDock's data is useless for training, as the algorithm it uses, called Lamarkian, would only apply to the specific receptor being docked. This is because with the type of screening we do, we specify a search space on the receptor for the ligands to hopefully bind to.
>Everybody in big pharma does so.
Any links or papers? I would be interested in reading about how they go about doing this.
>>53564387
>I think true AI is a still a long way off
How do you empirically verify that a program running on some hardware is "True AI" ?
Serious question.
>>53561981
We LITERALLY just made an AI capable of beating top level professional players in Go, a game regarded as the last traditional game that computers couldn't beat humans in. Stop talking out of your ass.
>>53562721
If we're just trying to replicate the human brain why not just have more babies?
>>53564441
>Any links or papers? I would be interested in reading about how they go about doing this.
I'm not a specialist at all, but http://www.ncbi.nlm.nih.gov/pubmed/25724101 and etc https://www.google.com/search?q=machine+learning+in+docking
>AutoDock's data is useless for training, as the algorithm it uses, called Lamarkian, would only apply to the specific receptor being docked. This is because with the type of screening we do, we specify a search space on the receptor for the ligands to hopefully bind to.
Looks like Lamarkian is just genetic algorithm (random search). Still, the idea of almost every ML approach in pharma is to save results of lots of costly docking simulations for various receptors/ligands to database, then train a simplified ML model that can output a good enough estimates of parameters of interest for some arbitrary receptor/ligand pair.
>>53561981
>what are expert systems
>what is a search engine
>what are modern gas cars
>what is a videogame
>what is machine learning
>what is basic database management
>what is love
>baby don't hurt me
>don't hurt me
>no more
>>53564460
>impressive, but still not AI
>>53564543
>what is love
2
>>53564502
This is an interesting paper, and I'm going to download it and read it tomorrow. Thanks anon.
>>53564549
>not ai
I don't think you know what is AI
>>53564614
Have it start with "no prior knowledge" and learn how to fly. Have it start with no prior knowledge and learn to speak... etc.
>>53564645
I know what you think AI is. I think that is just pattern matching. Impressive, but not intelligence.
>>53564666
That's better, but still there is a need in precise definition of experiment. Deepminds virtual experiments are just like that - the agent is put into the virtual environment, and it learns to accomplish a task on its own.
>>53563942
do computational neuroscience if ur school has that
but don't expect to get a job after graduating (you need at least a master's)
source: i graduated last year and do shit that is totally unrelated to what i studied in school
>>53564679
You are thinking about computers with self conscience?
If you develop a tic-tac-toe game where you can play against the computer, than this game has artificial inteligence.
And about pattern matching, of course they used a algorithm to learn what is good and bad in the game, just like we, humans, do.
>>53564693
I'm sorry if I offended anon. I'm sure your work is impressive. I just think AI has been going down the wrong path for some time. Maybe your deep learning project is the way to go. I'm not in this field. Just some opinions from an old guy.
>>53561981
>>53562034
>>53562323
>>53562643
https://youtu.be/fvtrRGmv7aU
>>53564764
Not him but I personally think we just have to develop more complex and appropriate systems, and the only way for our dumb brains to stumble on this may be through developing "AI" like ones that beat games like Go to develop frameworks that will serve as PARTS of true AI.
>>53563327
I think you also need Differential Equations brah
I think mainstream ai researchers are doing it wrong. The design of the neural networks doesn't have to be complex, it just has to be functional. What I propose is just training a bunch of easy to make easy to train feed forward networks to mimick functions of a regular brain, and piece them together in a way that allows them to send outputs and inputs to each other much like thoughts.I think is that there would also need to be a brainstem-like nueralnetwork that acts as a gateway to the final outputs, as in the outputs that would guide the actions of the desktop assistant or robot or whatever.
Why does everyone here need a human like machine anyways? We have humans for that. It is better to make machines that excel at particular things.
>>53564839
They aren't strictly necessary for deep learning. The only equation is gradient descent update, or its variants.
>>53564859
This. We can grow natural neural networks for the purpose of artificial intelligence. Making it with silicon is pointless (unless you wish to be able to upload your own consciousness someday)
>This thread
>>53564850
There is skip-thoughts approach and there are highway networks with skip connections that are similar to your ideas.
Separate training isn't encouraged though, people like end-to-end learning because it achieves best performance by optimizing the system as a whole.
>>53564859
To have them work for us, anon! Building one universal machine can be easier than building a ton of specialized ones. Honestly, I don't want to work. The sooner we free our fellow humans from mental and physical labor, the better.
>>53564859
A numerical simulation would allow for developing infinite analytical solutions to do application-specific tasks.
>>53564911
b-but I have these weights, they are made by my own implementation..
>>53564912
Child labour, third world countries, etc. Is that enough work force for you?
>>53564764
Biological path is cool too. We'll see how it goes!
>>53564968
Not my cup of tea. I'd rather have noone working. Why work? There is no law of nature that says that organism should do "work" to stay alive. Its just that our tech and econ aren't up to the task of freeing us (yet).
>>53564956
step it up senpai
>>53565016
Yup, time to upgrade my network! I'll do my best!
>>53565037
have fun
fuck, if there's more ai/nn threads on /g/ maybe i'll be motivated enough to finish my dqn[1] japanese rhythm game bot
[1] >>53562643
>>53564956
That's <20 lines in Tensorflow
https://www.tensorflow.org/versions/master/tutorials/mnist/beginners/index.html#mnist-for-ml-beginners
>>53564859
For fucking and seeing them climb stairs, of course.
But otherwise only linguistics really cares about acting like humans instead of just whatever gets shit done.
>>53565113
That's too easy. I did mine totally from scratch, without any foreign libraries. I will add convolutions and RNN in the same fashion. Tensorflow is cool, but I like to get to the basics.
With large architectures I'll need tensorflow though.
>>53564493
Because that's boring. Also
>implying any of us will ever procreate
>>53561480
I've read a few books on it (specifically artificial neural networks). It's more of just another tool in the toolkit to solve certain types of problems. Interesting stuff though, wish I had more time to read about it.
>>53563968
>>53564070
>>53564724
Thanks all of you
I guess i'll be doing it the way i planned
See you in the Singularity!
>>53562287
Do you have any sites/guides/tutorials where I can start on some ML?
>>53568297
http://scikit-learn.org/stable/tutorial/basic/tutorial.html
https://www.tensorflow.org/versions/0.6.0/tutorials/mnist/beginners/index.html
https://www.toptal.com/machine-learning/machine-learning-theory-an-introductory-primer
