The art of neural networks | Mike Tyka | TEDxTUM

Did you know that art and technology can produce fascinating results when combined? Mike Tyka, who is both artist and computer scientist, talks about the power of neural networks. These algorithms are capable to transform computers into artists that can generate breathtaking paintings, music and even poetry.
Dr. Mike Tyka studied Biochemistry and Biotechnology at the University of Bristol. He obtained his Ph.D. in Biophysics in 2007 and went on to work as a research fellow at the University of Washington, studying the structure and dynamics of protein molecules. In particular, he has been interested in protein folding and has been writing computer simulation software to better understand this fascinating process.
In 2009, Mike and a team of artists created Groovik’s Cube, a 35 feet tall, functional, multi-player Rubik’s cube. Since then, he co-founded ATLSpace, an artist studio in Seattle and has been creating metal and glass sculptures of protein molecules. In 2013 Mike went to Google to study neural networks, both artificial and natural. This work naturally spilled over to his artistic interests, exploring the possibilities of artificial neural networks for creating art.
This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at ted.com/tedx

[@hypercoder6577]

This deserves way more recognition by the general public...

[@tharindawicky]

Generative neural network images are amazing , just like we are dreaming... wow

[@radsimu]

you could input the image of a human and ask the network to generate the closest version of that picture that would be clasiffied as a dog. This way you could see the dog version of you. This is amazing

[@Re-workCo]

Amazing how results looked when Mike started with random noise and kept zooming in and the network keeps over interpreting the image!

[@anamp.chauhan1510]

Great job. Quite informative,story telling was quite good.

[@AbhishekButola]

i had made a project using neural networks. it was similar to detection and classification.. but i didnt knew we can reverse this thing too.. great concept!!!!

[@OnTheBrinkBook]

Okay I just have to pause that and say this is the trippiest picture I have ever seen.

[@user-jy5tc3mo4i]

This amazing study opens up some of the mystery of how our brains work

[@aritramishra685]

Absolutely loved this talk.

[@MrArunavadatta]

The concept of generative artificial neural network is very intriguing..

[@rzaaeeff]

I'm going to take the project about neural networks, which my teacher has been offering to me since last year :))

[@assaulth3ro911]

Oh my goodness I wish someone offered this to me!

[@rzaaeeff]

It seemed so complex and hard. That's why i didn't accept. After writing this comment, i contacted him, he told me it's not available anymore :(

[@tatanpoker09]

I was scared of learning neural networks first, so I forced myself onto writing an essay about it, which forces me on researching everything first and making one on my own. Still on the process, my brain melted around 2 weeks ago but it's slowly getting stuff together

[@christianpaoliello2826]

I know that feeling. It took me some months to go from 0 to build a NN that can classify single digit numbers. My main resources were the free Machine Learning course of Andrew Ng, and an old book "Fundamentals of Neural Networks - Fausett" which I think to be good if your math knowledge is not advanced and want a practical book with pseudo-code. Because I was in highschool, I could not handle the math that now I believe to be fundamental in order to fully understand NNs.

[@MrLudiJasko]

Hey Christian! How is your progress nowadays with ML? Are you still involved in this area?

[@LexPodgorny]

Perhaps, the most inspiring video I have seen on youtube.

[@TheAIEpiphany]

Awesome stuff I love the interaction between the technology and the art!

[@nikitaseverinov]

impressive work methodology of neural networks for images, thanks!

[@hugel99]

The images by single neurons are fascinating.... good job!

[@Michaelangelo1000000]

The most interesting thing for me is effect in neural networks, which can be considered as music. Set of sequences in neural network, leading system to achievement of “goal”, is algorithm. Human music is algorithmic sequence with weight functions, which can be shared and make thousands of people do the same things (pendulum effect).

[@riquardi]

Interacting with this video will increase the likelihood to get more like this in my feed :)

[@AMITSHARMA-oh1nq]

At 12:47 ...... I feel the beginning of something incredible, exciting,wonderful era where machines thinks and do osame works that human does

[@Mornys]

It's like intuition, but it can be fed more and different type of data a human can take as an input and do it much faster. Then it just blurts out the answer it's trained to do without being able to deduce if it's right or not. Still very powerful and can take us huge step forwards on many different disciplines. Very interesting.

[@shreeniwasyesare]

Nice presentation and interesting examples to understand Neural networks.

[@akshayarekhasubramaniyan9537]

some data stored in today!! got my neurons fired up!! thank you sir...

[@gmarciani]

Amazing #neuralart and great presentation! thank you

[@mounica2035]

loved the presentation

[@IvanReyes111]

wow, awesome presentation.

[@anntsh1031]

Wow that was so impresing!!!

[@sudarev_gleb]

Интересная презентация и крутая подача простыми словами на тему «что такое нейронные сети». Единственное, я бы больше углубился в технические особенности, то есть, как именно происходит сие создание. Иногда кажется, что это настолько тяжело, что даже самые великие умы не могут создание нейросетей описать одним-двумя предложениями.

[@gmshadowtraders]

Very interesting presentation!

[@radsimu]

this is very trippy while sober

[@monkeyrobotsinc.9875]

the fractal art stuff is incredible.

[@insoaryt]

Hey Mike! You expected people to applause you after novel generated by nn, just like after a music composition. But they didn't. Can you guess why? I could guess. Just because they were in such a shock. The best people can do will be soon overwhelmed by machines... some smile here :-) Don't know why, but I smile, perhaps because as I think my guess was right. Just wanted to share it here. P.S.: And yep, I applause!

[@rogeliomoisescastaneda7396]

Amazing!

[@edselmalasig]

thanks a lot mike

[@eduardoxmenezes]

beautiful

[@chromaspark1257]

The images it generates make me really uncomfortable for some reason.

[@arooobine]

9:20 OMG I'm watching a neural network dream!!!

[@Jamiefearon]

No, you're watching an artificial neural network experiencing the equivalent of LSD

[@cabmok]

You mean nightmare!

[@MrYonikeren]

at 8:27 he said that generative nets started by experiments by Alexander from Google. Can anyone please help me find that resource? I've been looking for it, but couldn't find it.

[@AlbaXhani]

research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html Does this help?

[@MrYonikeren]

Oh yes!! Thanks Alba! :)

[@sebastianvatne1640]

Anyone know what the names of the computers he is talking about is? At 4:26.

[@elfboi523]

I wonder whether you could use a neural network to create industrial music as well, by training it on all kinds of rhythmic, repetitive machine sounds.

[@plavix221]

+elfboi523 of course.......... They will create everything..... virtual worlds.........basicallyyy...........everything........... this just the tip of the iceberg what we are seeing.... last week AI created new organic materials.....

[@gcgrabodan]

+MrMinusguy if your "results", ie. music, or your code are online please let us know ;)

[@gcgrabodan]

+MrMinusguy or at least let me know, who cares about the other 6.9 billion idiots :P

[@elfboi523]

***** I was thinking about something that would sound more like early Einstürzende Neubauten.

[@salemalgharbi8834]

This Ted presentation explains clearly Stephen Hawking says:” The rise of powerful AI will be either the best or the worst thing ever to happen to humanity. We do not know which”

[@bobdobbs7795]

Exactly: the blessing AND the curse! It just depends on the dosage, timing, and intent.

[@ifacarts]

Right on

[@akashshrestha01]

Awesome

[@sambowwow21]

can someone please help with the title of Turing's paper on computational architecture ?

[@jabberwokiuk]

+Sami Yousif, You may have already found this by now, but you can find reference to it in The Essential Turing : books.google.co.uk/books?id=VlC5MkVIwqkC&pg=PA422&lpg=PA422&dq=organising+unorganised+machinery&source=bl&ots=5ghWjyIDlf&sig=bQKMkGH7f9RXJ9N8Ymrp-o82_SU&hl=en&sa=X&ved=0ahUKEwidgdDWsJrMAhXLECwKHc_kBH4Q6AEIIjAB#v=onepage&q=organising%20unorganised%20machinery&f=false

[@marinellovragovic1207]

1:09 Who thought of Life is strange, after seeing his name? 😁

[@askformoreinfowhichyouwont7510]

Where did he go to study neural networks?

[@ShirishJadav162]

lower laryer detect abstraction while uper layer detects whole pictures I have a arch in mind that can be used to make a core of ANN such as GPU. but its not digital its analog and would be faster in Training and faster and parallel in operations as compared to Digital Processing. I would document it up and send you.

[@darksannhius]

+Shirish Jadav Why would it be faster? Using digital or analog units doesn't change the underlying algorithmes, and it is those algorithmes that need to be improved, and it is improving (for example, batch normalization brought a good boost to learning speed not a long time ago), this area is moving incredibly fast, and i don't think any analog setup would beat a good gpu (it's like saying "yeah i have this idea for a steam plane that would be faster than most planes!", sorry for being this ruthless).

[@ShirishJadav162]

yes but digital signals need clock ticks .. while analog don't still both systems have delay.. bt in case of digital the delay is more.. there are some other things analog system have problems with like environmental noise etc. and it analog design is a lot harder process.. bt if done in analog domain we can have parallelism using different frequencies etc.. lot of things are possible with digital systems can't do.. bt some how finding a balance in between hardware implementation of both of these is required. but lets see where it takes us. I completely understand what you are thinking.. but you see most of us have just seen the digital systems... but if you actually study analog domain you would understand it... btw there are research papers on designing neural networks with analog design. and the think is I first thought the idea is new bt no people have done work on this idea bt believe me.. its difficult bt its what happens in brains and real life or nature, nature is not digital...

[@ShirishJadav162]

there have been many analog computers.. search analog neural networks. second is don't be biased it may be I am wrong(which is true in current scenario) and you may be false in future.. :P :) so its ok you don't see it.. I do see it. my idea is just a concept right now.. :) If given resources I can prove that its possible and I can do it.

[@massimookissed1023]

+Shirish Jadav , biological NNs are analogue, digital ANNs attempt to model the biological ones. Analogue multiplication and summation are instantaneous, digital requires discrete clock ticks or at least propagation delays.

[@ShirishJadav162]

Right so if we focus on analog which I think in some remote part of world some secret R&D would be doing... ;)

[@reyanshmishra7676]

Wow!!!!

[@aleksgorskov]

these systems will be very helpful ATMs around the corner. but it's still many many years until they will be able to converse the way we humans do. because when I see a picture of a tree I say "it's a tree" bat if shown a tree midst of field of crops I may suggest "its sad to be left alone"..

[@theodorechandra8450]

that music.......

[@AshtonSnapp]

Make a Neural Network write a book!

[@theropoy9371]

It can

[@TheFatblob25]

This is so interesting...order out of chaos. A unique expression of higher level consciousness & technology & yet In some ways "The Great Mediocre" .

[@massimookissed1023]

If it has a hub-cap, it's more likely to be a dog than a cat. Cats don't chase cars so much.

[@ShreyasGameboy]

so that's how Prisma works 8:40

[@kittykatBflat]

I'm thoroughly creeped out.

[@Jamiefearon]

"an untrained neural network classifying cats and dogs when shown images of cats will get it wrong most of the time" No, wrong! It will get it wrong 50% of the time. Only a highly trained network could get it wrong (or right) most of the time.

[@GameCarpenter]

If the network is more or less as diagrammed, 50% success is the expected value, but not a reasonable assumption of a specific result. As far as the statements go, I suppose it's dependent on the interpretation of "most." Using the easiest mathematical interpretation of more than half, the statement in the video (assuming the quote is correct) is more than half wrong in one particular way. You would expect that an untrained neural network would have less than 50% accuracy as often as it would have more than 50% accuracy, and then sometimes, it would have 50% accuracy. JF's statement in his comment about highly trained networks with this interpretation of most would be almost infinitely wrong, since it would only be right if an untrained neural network were exactly 50% correct before training. A more general interpretation of most, however, would tend toward a higher percentage. As the interpretation of the meaning of "most" moves toward 100%, the video comment becomes inaccurate for a different reason; the tendency for a sequence of random events to cancel each other out, make it increasingly unusual for a random neural network to be wrong "most" of the time. If being right "most" of the time were sufficient, you could then just flip the outputs and be done without training. With this interpretation, JF's comment about highly trained networks simultaneously becomes more and more accurate. Either way, it seems like that line from the video is at least a slight misstatement. The idea though is of course, to convey that an untrained neural network is pretty terrible at predicting things.

[@user-ol2gx6of4g]

Simon WoodburyForget You don't know what you are talking about.

[@stephenborntrager6542]

It can be wrong *most* of the time; typicallt, you would output more than just cats and dogs. Its not a 50/50 chance. A particular implementation of tensor flow I know of is wrong 90% of the time, often mistaking humans for refridgerators, or french horns.

[@leo333333able]

cool

[@ShirishJadav162]

are there libraries in Python for deep neural network Experiments??

[@huluvublue112]

+Shirish Jadav 'Pybrain' is a great library for that. Incidentally it was developed at TUM (I love my uni ^^)

[@ShirishJadav162]

+Huluvu Blue thanks for that I would look on that . I hope its documented too. ;-) well I am designing a hardware that could be faster and more like real neurons so before that I need to by hard the math .. because I have basics bt its always math of setting weights that troubles me. :-)

[@jake3189]

+Shirish Jadav tensorflow

[@beegieb4210]

+Shirish Jadav Tensorflow and Theano are your best bet. If you want a little more abstraction, go for Keras or Lasagne. Don't use Pybrain, it's horrible compared to modern libraries.

[@noelearlwatson2724]

Sunspring

[@kpbuzz]

mind = blown

[@nunobartolo2908]

just like dreaming

[@mihir469]

One question. Is this just a decision tree???? DONT LIE

[@musikSkool]

What if you teach it a face at all 180 degrees (of the front). Then reverse the process and ask it what it thinks the face looks like at a particular angle? (I just described computer graphics in 2030, btw.)

[@martinsalko1]

that's what's great and also creepy about them. you have input then magic then output. and you can try your best but you'll never know wtf is happening in there.

[@VVe11erMichae1]

But it really isn't magic, it can all be explained and understood. Try understanding Convolutional neural network, they show how its possible to represent an image by its features.

[@suvtropics]

O: