When we're not accidentally creating abominations we're accidentally creating computers who create abominations. GG. No, but seriously, this is an amazing development and I'm quite intrigued as to how it will progress over time.
IceMetalPunk by that logic; youre saying that the computer also has an imagination. Because it took a %100 random image and created an image using (basic) pattern recognition.
Saltsugarsex C. pretty much exactly. and that is some serious imagination. computers have better imaginations than people confirmed. Now even artists jobs are not safe from machines.
Saltsugarsex C. Yep. This is an example of an artificial imagination. It's limited only by the categories it learned about during training, but that's true of humans; we just have many years of constantly learning how to categorize everything whereas this system only had a relatively short time with a relatively small search space (i.e. input size, i.e. "things to learn about").
IceMetalPunk As a recent computer science graduate who did an honours seminar on convolutional neural networks, I can tell you that even though this may look something like human creativity or imagination, it's not even remotely close. In ANNs, a structure of simple processing nodes (sometimes even a simple threshold/sum comparison) is recognizing patterns using sequentially more detailed scans - or at least this is the easiest way to think of it. I'd do a bit of reading up on how deep learning neural networks actually classify objects before thinking that a neural network can exhibit anything like the amazing emergent property of our brains that is imagination! They're orders of magnitudes apart in terms of complexity, and function in completely different ways. You won't be seeing an ANN replacing an artist any time soon, believe me :P
ghudner I'm also a computer science grad. I never said they work the same way as a human brain, but that doesn't mean they're not imagination. If you zoom out of the detailed implementations and look at the abstract concepts, these are generating new images by combining features of old ones based on learned pattern recognition. Which is exactly what imagination is. It doesn't have to be implemented the same way as a human brain to be imagination. Kind of the same way DFS and BFS are different, but they're both still search algorithms that can find a node in any given connected graph.
TheJaredtheJaredlong He has lack of data about himself, and holds no self image. And there's no picture for a program, because a program is a digital thing. Basicly he will fail to make a picture of himself.
TheJaredtheJaredlong That's what's already happening, these images were created by passing them through googles algorithm creating a new slightly different images then passing it through again and again.
Rylee Matthew well if an advanced AI in future come up with this nightmare fuel, I wonder if it would be terrified if it saw things the way we see them (provided, of course, that said AI is equipped with the ability to feel fear)
+Ferman Sensei At the same time, the pattern seeking/interpreting process has some striking similarities to how our minds do it. Primitive, yes, but only for now. Maybe, in the process we can learn about ourselves. Essentially, we both run on electricity making specific connections within a vast network of hardware (organs, nerves) and software (DNA, genes) sending and receiving instructions.
What would be more disturbing would be to let that daydream computer have access to the internet, as in general http browsing capabilities and let it teach itself what an object looks like. After all, people tag EVERYTHING. And if the program can become complex enough, it could probably create forum threads asking to identify what's in that image... Well, that might be 1-2 years in the future from this point, but we're actually eerily close to real AI.
d3rrial These kind of networks only start to work after they've been fed thousands and thousands of pre-determined training data. So in order for it to know what a bird is, the designer has to have thousands of pictures of birds to run through the network learning program before the network can effectively identify birds. That also means that these networks become optimized to certain tasks. Networks that identify cars won't be good at identifying cancer cells (there are networks that identify cancer cells, and they often do it better than humans). Unfortunately, Neural Networks are still a good ways off from true AI capabilities. Neural Networks are fantastic at sensory type problems. They can identify images and sounds better than humans, and they're used for speech recognition and translation technology. And they accomplish this through large amounts of mathematical functions. A network isn't thinking, so much as it's doing lots and lots of optimized calculations. So while we are close to having vehicles that know how to drive themselves, we haven't yet figured out a good way to make a program engineer a machine, perform scientific research , create new fields of mathematics, or program themselves. We'll most likely enter a time when humans and programs interact in a closer manner and solve problems together. Until we can design a system of higher level thinking, AI won't be able to take over the reigns of civilization.
d3rrial I wouldn't recommend giving this kind of algorithm access to the whole Internet. Just download a couple terabytes of data from Shutterstock and Facebook and let in run wild with that, first. But yes, this sounds like the "final approach" for AI. What an interesting time we live in!
thepokkanome What you call "thinking as human", such as "engineer a machine, perform scientific research , create new fields of mathematics, or program themselves", is what people call creative activity, and what actually is (from mathematical point of view) solving of an NP-problem. Majority of humans solve majority of NP-problems much slower than computers (e.g. sudoku). Only thing human brain has priority in is a technical part - brain is much more space- and energy-efficient and organized than computer. There is nothing supernatural in human abilities to invent staff and write music, draw pictures and write books, it's just product of a huge amount of information absorbed and randomly combined. While this google program was fed with only 1,2 mln pictures, tagged with words, human brain consumes insane amount of pictures with sound, words (things in itself) and smells for up to 100 years. What I wanna say, is that what you disparagingly call "not thinking, but optimized calculations" is much more effective to solve NP-problems than human brain's "randomly mix up everything and send to output if ok" (disorganized, often not-optimized calculations).
Егор Свежинцев I have to disagree. Computer programs, as we currently write them, follow a defined pattern. Humans are still much more efficient at figuring out new patterns to come up with creative solutions compared to a computer. It's not impossible for a computer to that kind of problem solving, but our current methods take a lot of time and computing power. So even though computers are already better at sudoku, and they can beat us at chess or jeopardy, it still took a human developing an algorithm to get the computer to accomplish those tasks. Humans are still better at solving new NP-problems. We don't have computers that write algorithms yet (at least I've never seen any). Eventually we'll get there, but that's roughly what I meant by "not thinking".
Thank you so much for not saying that it works like the brain. It's a common mistake people make when talking about neural networks. The original ones were biologically inspired, but never were very biologically plausible and have evolved so much over the decades that they no longer resemble the brain at all.
InnovumTechnology Well, there actually are atempts to make artificial versions of biological neural networks (the blue brain project for instance), but yeah, they are not artificial neural networks in the normal sense.
InnovumTechnology oh, really? I was under the impression that these neural networking method were how the brain processes information. Obviously not using neurons or anything like that, but with similar processing to allow for similar outcomes.
RPGgrenade They aren't. The only things they have in common are that there are neurons and connections of varying strength between them. There is a lot that differs. For example, backpropagation is a very common feature of neural networks, but is definitely not occuring in the brain. From there, synapses in the brain are a lot more numerous than those in neural networks, and behave very differently. They are much less reliable. Neural networks are essentially like a combination of regression analysis and a taylor series. It's basically taking an enormous amount of data and creating a set of functions that, when added together, give you a function that fits the data reasonably well. It's a little more complex than that, but that's the basic idea. The problem is that they require an enormous amount of data to be practical and need separate training and operational stages. In other words, they don't learn continuously, and they only learn things after seeing them a large number of times. It can't be said for certain how the brain works at this point, but the behavioral properties that the brain has are clearly very different from those found in neural networks. The most biologically plausible theoretical model that I know of is Jeff Hawkins' Hierarchical Temporal Memory. It's based on known behavior of dendrites, synapses, and cortical columns (organized structures of neurons in the cortex) to create a system that models incoming information pretty well. The only real problems with it at the moment is trying to get information out of it, and trying to get it to generate behavior.
Does anyone know if there is research being done about the huge similarities between this and psychedelics like LSD or Psilocybin mushrooms? The images you get from said psychedelics can be extremely similar to these, so I guess we could learn something about our brain and how it recognizes and replicates patterns. Which part of the brain is responsible for this and how does it work, etc.
Mussi93 Looking at those images, the first thought that occurred was that they look surprisingly like stuff I've seen when taking psychedelics. Interpenetrating this from my own subjective experience, the similarities make me feel that this algorithm may be quite similar to part of the human minds function. Such a fascinating topic, I'd love to see more reporting on scientific research in this area, I can't get enough. The fractal patterns one sees under the influence of psychedelics must say something pretty important about how our mind works, just wish I could understand it more.
Mussi93 My intuition said's that the image you see on psychedelic drugs are the result of the pattern recognition systems in your brain going into overdrive, what should usually be filtered out or never recognized at all is being sent to the lateral areas of your brain instead. It would explain why things like faces are the most common pattern recognized while on psychedelics as that is the thing your brain put's the most effort into looking for.
This is cool :) I think a way to improve would be to select the images from the image bank, say 100 of this type, 100 of another type, so there's not loads of dogs.
I've done some work in my career (and education) with neural networks. There's one error in this video: in general, you don't manually modify the nodes of a neural network. Instead you use one of many back-propagation techniques during the training process. The input to the network is an unknown image and the output of the network is "back driven" with the solution. This process modifies the weighting of the various nodes and their connections between and inside layers. The engineer of the neural network might modify the number of layers, number of neurons in a layer, or even the types of neurons in a layer but they would never tamper with the nodes and their weighting algorithms directly. I wrote a neural network in MatLab and trained it to recognize distorted letters (like in a captcha, but it was for old books with missing ink and water stained pages). The crux of it was that the training included human identification. So the neural network was as good as a person at identifying distorted characters, but no better.
It is humbling to know that no matter how much you teach computers with all of their technological wonder, the human brain is still infinitely more powerful!
Crazy to watch this now. We went from this to Stable Diffusion in 7 years. It's crazy how this technology came out of nowhere. it wasn't even science fiction a decade ago, it was science fantasy. the idea of teaching an AI how to intuit and understand vague concepts? it's unbelievable.
I didn't want to watch the video now but I did because I was curious now I'll have nightmares thank you google also(1) mildy disturbing. "mildly"(!!!) also (2) I feel like this deepdream computer makes images like toddlers talk. you know when toddlers learn to talk and they say the words all messed up with like 10 extra letters... kinda like that.
This is all rather creepy to me. When I would try to fall asleep when I was a kid I would see images like those when I closed my eyes. I've since grown out of it, but seeing these things again makes my skin crawl. Thanks, Google!
It's ability to find the "dog" in a picture intrigues me, because it seems to open up possibilities to use computers to analyze and find patterns even more precisely than humans do. Just like humans could always draw curves, whole new possibilities opened up when we programmed computers to do that more perfectly. For instance we never be able to analyze the paths of particles in 3d space after a collision, we need to program computers to do that accurately. Similarly, there is probably something new we could get computers to do more accurately and precisely in the realm of previously holy human grounds of abstraction.
Speaking of recognizing things, I was thinking about this recently: How do animals recognize objects? For example, let's say a human is walking outdoors with a walking stick, and an animal spots it in the distance. Does the animal recognize a bipedal animal with two arms that's holding a stick, or does the animal recognize a three-legged creature? Does it assume the stick is part of the human, or does it know that the stick is separate from the creature it's looking at?
Even Stephan Hawking thinks "skynet" is a possibility. This is 1 step closer to it, this of a computer that can recognise what humans deem as safe or good hiding spots.
What about sentient Microwaves? Constantly judging you on your eating habits and asking if you know any single microwaves wanting to link GUI's for a computer night.
2045-2090 according to the experts. The later they arrive, the more work we will have put into friendly AI though, and the less likely it is to use our atoms for something else. Though if you tell more people about Roko's Basalisk, those times should drop a bit.
This reminds me of when I was little and when I would stare at the patterns on a tile floor looking for the group of splotches that looked the most like a face.
Really interesting stuff. I'm interested in how we're actually working in creating smarter AI. Feel free to make more videos about this topic, and maybe some references to movies like Ex Machina or even Chappie.
Like IceMetalPunk said, it's quite possibly just like imagination is side effect of human pattern recognition abilities. But it's also quite profound. Even if this doesnt lead to A.I directly (although it could), the fact that researchers already need to probe their system to understand how it works means that, once A.I emerges, people will have a hard time realizing that it had.
Vaidas Šukauskas A lot of technologies were invented, but took many years before they were fully integrated into society. With how the Internet is approaching a sort of "cultural singularity", and how increasingly interconnected we are, I wouldn't be surprised if AI *is* developed but rather than changing everything all at once, it gradually seeps into our lives. The philosophical questions are asked one at a time.
That sort of situation, where you don't know how a piece of software works until you probe it, is somewhat unique to all learning, machine or otherwise. When a human learns, synapses form and rearrange in the brain, and we can't yet look at those structures and immediately infer how they work (though we continually make steps toward that goal). We need to ask them what they're doing--"probe" them. The same is true here: the neural net is constantly adjusting its weights, but to humans, those are just a bunch of numbers. It's extremely difficult to look at the organization of numbers and directly infer how the network is running. So instead, we "ask" it to walk us through its process with an output at each step, the same way we'd ask a child how they worked out their homework exercises. When general AI is born, we will know immediately, because it's defined not by its methodology, but by its behavior. And the behavior of a system is far more visible to us than the processes that lead to said behavior.
What I find most disturbing about google's daydreaming software, is many of them make me extremely uncomfortable with how close they are to a couple... college experiences. I didn't even know it was possible to portray that in an image...
+Richard Smith ever hear of Alex Grey ? great artist , definitely has done ayahuasca ceremonies in which one travels with the use of DMT from plants in the Amazon rain forests .. it is a naturally produced chemical stored in the brain and released at birth and death ( as well as near death ie: stories of light at the end of a tunnel ) .. the long and short is that the images produced by deep dream are quite interestingly more than similar to what one sees during such a journey
I really hope they come out with different publically-accessible variants of Deep Dream designed specifically for detecting and enhancing different patterns than just dogs and eyes for such things to produce different aesthetics from the photographic enhancement. Like, an industrial one designed for enhancing machines & tools and smoke in things, a horror one designed specifically to enhance blood; gore; skulls & monsters, one perverse one designed for enhancing furries; tentacles & genitalia because god is officially dead and I am trash. Would be neat from an aesthetic standpoint is all i'm saying.
*What everyone sees:* A robot imagining something artistic and creative. *What it really is:* Random noise produced by a confused machine that has no clue what it's even doing.
Thes images resembles those Hypnagogic Halucinations; when you first open your eyes after waking up, and you see objects transfigurated into something else for one or two seconds.
My own dreams resemble these more than it does inception... But at least my dreams make aesthetic sense. These images don't, but they are very enjoyable in a vicarious drug trip sort of way.
Will you go over the science of the Sony project N prototype (parabolic speakers emitting sounds to your meatuses while not needing ear phones in your...meatuses)???? Also, something to do with voice recognition, like how a computer converts analogue voice or sound into binary through the use of sampling, and compression/decompression algorithms??? Maybe how Microsoft Kinect works??? All sound good.
Ha, I gave a big post on one of scishow's videos explaining neural networks before. More in the context of game-playing AI, though, rather than image recognition.
Nice demonstration how to _project_ meaning where there is none. At this rate, we might have the first robot religion on our hands in the next 20 years.
When we're not accidentally creating abominations we're accidentally creating computers who create abominations. GG.
No, but seriously, this is an amazing development and I'm quite intrigued as to how it will progress over time.
Zetsubou Z. Wat..?
***** Found the idiot guys.
Leave it to Monokuma to be unimpressed by DeepDream.
Kappa
Zetsubou Z. I mean let's be real, we're creating computers on purpose that create abominations. :P
SoosUnknown "When we're not accidentally creating abominations "
Yeah, I am an idiot for not understanding what he thought when he wrote this..
I for one welcome our psychedelic overlords.
01001000 01100001 01110010 01100101 00100000 01001011 11100001 10111001 10011011 11100001 10111001 10100011 11100001 10111001 10000111 01100001 00100000 01001000 01100001 01110010 01100101 00100000 01001011 11100001 10111001 10011011 11100001 10111001 10100011 11100001 10111001 10000111 01100001 00001101 00001010 01001011 11100001 10111001 10011011 11100001 10111001 10100011 11100001 10111001 10000111 01100001 00100000 01001011 11100001 10111001 10011011 11100001 10111001 10100011 11100001 10111001 10000111 01100001 00100000 01001000 01100001 01110010 01100101 00100000 01001000 01100001 01110010 01100101 00001101 00001010 01001000 01100001 01110010 01100101 00100000 01010010 11000100 10000001 01101101 01100001 00100000 01001000 01100001 01110010 01100101 00100000 01010010 11000100 10000001 01101101 01100001 00001101 00001010 01010010 11000100 10000001 01101101 01100001 00100000 01010010 11000100 10000001 01101101 01100001 00100000 01001000 01100001 01110010 01100101 00100000 01001000 01100001 01110010 01100101
Human imagination is a side effect of pattern recognition. It's no surprise, then, that computerized pattern recognition can also imagine things.
IceMetalPunk by that logic; youre saying that the computer also has an imagination. Because it took a %100 random image and created an image using (basic) pattern recognition.
Saltsugarsex C. pretty much exactly. and that is some serious imagination. computers have better imaginations than people confirmed.
Now even artists jobs are not safe from machines.
Saltsugarsex C.
Yep. This is an example of an artificial imagination. It's limited only by the categories it learned about during training, but that's true of humans; we just have many years of constantly learning how to categorize everything whereas this system only had a relatively short time with a relatively small search space (i.e. input size, i.e. "things to learn about").
IceMetalPunk As a recent computer science graduate who did an honours seminar on convolutional neural networks, I can tell you that even though this may look something like human creativity or imagination, it's not even remotely close. In ANNs, a structure of simple processing nodes (sometimes even a simple threshold/sum comparison) is recognizing patterns using sequentially more detailed scans - or at least this is the easiest way to think of it.
I'd do a bit of reading up on how deep learning neural networks actually classify objects before thinking that a neural network can exhibit anything like the amazing emergent property of our brains that is imagination! They're orders of magnitudes apart in terms of complexity, and function in completely different ways. You won't be seeing an ANN replacing an artist any time soon, believe me :P
ghudner
I'm also a computer science grad. I never said they work the same way as a human brain, but that doesn't mean they're not imagination. If you zoom out of the detailed implementations and look at the abstract concepts, these are generating new images by combining features of old ones based on learned pattern recognition. Which is exactly what imagination is. It doesn't have to be implemented the same way as a human brain to be imagination.
Kind of the same way DFS and BFS are different, but they're both still search algorithms that can find a node in any given connected graph.
Those aren't electric sheep!
THOSE AREN'T ELECTRIC SHEEP AT ALL!!!
firebornliger
so, I guess that means the answer would be, "no."
THEY’RE DOGS!
So it can recognize dogs and bananas, but can it recognize dank memes?
Lol
Just feed it "internet" for ten minutes....
But can it recognize Class VI Dental Cavities?
(Gotta make them dentists lose their jobs)
Thomaster шлепанцы Watson has that covered for you.
Oscar Smith
Ah, yeah right... That existed too...
I bet by 2030 the whole world will be unemployed.
hahaha XDDDDD the dank maymay meme, very well done, saved. Truly a king of comedy XDDDDDD epic
I want to see what happens when you tell DeepDream to make a picture of DeepDream. What will the program identify as it's "self"?
TheJaredtheJaredlong the program encounters a paradox and crashes.
TheJaredtheJaredlong He has lack of data about himself, and holds no self image. And there's no picture for a program, because a program is a digital thing. Basicly he will fail to make a picture of himself.
FlyingJetpack1 He... come on now, even though this program is very smart, it doesn't deserve a pronoun reserved for living beings.
rippspeck But it does deserve a pronoun for being entitled... are you replying to the wrong comment?
TheJaredtheJaredlong That's what's already happening, these images were created by passing them through googles algorithm creating a new slightly different images then passing it through again and again.
Who knew a computer's mind is filled with morbid pictures of eyes, disembodied limbs, and horrifying mutated creatures?
This will give me nightmares.
i almost puked once.
Rylee Matthew well if an advanced AI in future come up with this nightmare fuel, I wonder if it would be terrified if it saw things the way we see them (provided, of course, that said AI is equipped with the ability to feel fear)
Computer, find 'Illuminati'
Dorian Scott Computer says NO.
Dorian Scott You Rang?
Loomi Naughty Shhhh... Don't reveal yourself... *not yet*...
Dorian Scott I read that with a Jean Luc Picard voice.
Find green triangles... ( ͡° ͜ʖ ͡°)
This demonstrates how hard it is to actually create an AI with human capabilities.
There are some that argue that there are very few "capable" humans.
Ferman Sensei Why AI? Better have sex xD it only took us nine months
+Ferman Sensei At the same time, the pattern seeking/interpreting process has some striking similarities to how our minds do it. Primitive, yes, but only for now. Maybe, in the process we can learn about ourselves. Essentially, we both run on electricity making specific connections within a vast network of hardware (organs, nerves) and software (DNA, genes) sending and receiving instructions.
You had no idea
They can dream of dogs, but how about Electric Sheep?
The Android Next Door Hmmmm techno lamb.
kagutsuchi969 I always thought it was technoland ......
Master of Mundus
Well that's where the techno lambs frolic.
kagutsuchi969 all lifes questions have been answered
Master of Mundus
You're welcome my child. Now venture forth and enlighten the world with the knowledge you have gained.
What would be more disturbing would be to let that daydream computer have access to the internet, as in general http browsing capabilities and let it teach itself what an object looks like. After all, people tag EVERYTHING. And if the program can become complex enough, it could probably create forum threads asking to identify what's in that image...
Well, that might be 1-2 years in the future from this point, but we're actually eerily close to real AI.
That is very very possible right now.
d3rrial These kind of networks only start to work after they've been fed thousands and thousands of pre-determined training data. So in order for it to know what a bird is, the designer has to have thousands of pictures of birds to run through the network learning program before the network can effectively identify birds. That also means that these networks become optimized to certain tasks. Networks that identify cars won't be good at identifying cancer cells (there are networks that identify cancer cells, and they often do it better than humans).
Unfortunately, Neural Networks are still a good ways off from true AI capabilities. Neural Networks are fantastic at sensory type problems. They can identify images and sounds better than humans, and they're used for speech recognition and translation technology. And they accomplish this through large amounts of mathematical functions. A network isn't thinking, so much as it's doing lots and lots of optimized calculations.
So while we are close to having vehicles that know how to drive themselves, we haven't yet figured out a good way to make a program engineer a machine, perform scientific research , create new fields of mathematics, or program themselves. We'll most likely enter a time when humans and programs interact in a closer manner and solve problems together. Until we can design a system of higher level thinking, AI won't be able to take over the reigns of civilization.
d3rrial I wouldn't recommend giving this kind of algorithm access to the whole Internet. Just download a couple terabytes of data from Shutterstock and Facebook and let in run wild with that, first.
But yes, this sounds like the "final approach" for AI. What an interesting time we live in!
thepokkanome What you call "thinking as human", such as "engineer a machine, perform scientific research , create new fields of mathematics, or program themselves", is what people call creative activity, and what actually is (from mathematical point of view) solving of an NP-problem. Majority of humans solve majority of NP-problems much slower than computers (e.g. sudoku). Only thing human brain has priority in is a technical part - brain is much more space- and energy-efficient and organized than computer. There is nothing supernatural in human abilities to invent staff and write music, draw pictures and write books, it's just product of a huge amount of information absorbed and randomly combined.
While this google program was fed with only 1,2 mln pictures, tagged with words, human brain consumes insane amount of pictures with sound, words (things in itself) and smells for up to 100 years.
What I wanna say, is that what you disparagingly call "not thinking, but optimized calculations" is much more effective to solve NP-problems than human brain's "randomly mix up everything and send to output if ok" (disorganized, often not-optimized calculations).
Егор Свежинцев I have to disagree. Computer programs, as we currently write them, follow a defined pattern. Humans are still much more efficient at figuring out new patterns to come up with creative solutions compared to a computer. It's not impossible for a computer to that kind of problem solving, but our current methods take a lot of time and computing power.
So even though computers are already better at sudoku, and they can beat us at chess or jeopardy, it still took a human developing an algorithm to get the computer to accomplish those tasks. Humans are still better at solving new NP-problems. We don't have computers that write algorithms yet (at least I've never seen any). Eventually we'll get there, but that's roughly what I meant by "not thinking".
0:59 the way you said, " layers " reminded me of Dr. Evil.
lol
Finite Atticus next week: sharks with laser beams attached to their frickin' heads.
zack42187 lol. Apple's newest hottest product to compete with daydreaming Google machines.
Finite Atticus One million layers!
Finite Atticus I don't know, I'm getting distinct visions of ogres...
Thank you so much for not saying that it works like the brain. It's a common mistake people make when talking about neural networks. The original ones were biologically inspired, but never were very biologically plausible and have evolved so much over the decades that they no longer resemble the brain at all.
InnovumTechnology Well, there actually are atempts to make artificial versions of biological neural networks (the blue brain project for instance), but yeah, they are not artificial neural networks in the normal sense.
neuron1618
I know that. There is also Jeff Hawkins' work, which I think may be pretty useful if they can get it to generate behavior.
InnovumTechnology oh, really? I was under the impression that these neural networking method were how the brain processes information. Obviously not using neurons or anything like that, but with similar processing to allow for similar outcomes.
RPGgrenade
They aren't. The only things they have in common are that there are neurons and connections of varying strength between them. There is a lot that differs. For example, backpropagation is a very common feature of neural networks, but is definitely not occuring in the brain. From there, synapses in the brain are a lot more numerous than those in neural networks, and behave very differently. They are much less reliable.
Neural networks are essentially like a combination of regression analysis and a taylor series. It's basically taking an enormous amount of data and creating a set of functions that, when added together, give you a function that fits the data reasonably well. It's a little more complex than that, but that's the basic idea. The problem is that they require an enormous amount of data to be practical and need separate training and operational stages. In other words, they don't learn continuously, and they only learn things after seeing them a large number of times.
It can't be said for certain how the brain works at this point, but the behavioral properties that the brain has are clearly very different from those found in neural networks. The most biologically plausible theoretical model that I know of is Jeff Hawkins' Hierarchical Temporal Memory. It's based on known behavior of dendrites, synapses, and cortical columns (organized structures of neurons in the cortex) to create a system that models incoming information pretty well. The only real problems with it at the moment is trying to get information out of it, and trying to get it to generate behavior.
Oh look, pedantry.
What a daydreaming computer
and i thougt computers werent sloppy like me.
A lot of the pictures are what an acid trip looks like..
Really accurate with the shapes and colours.
Yes, install the code on your computer. That's just what Skynet wants.
Does anyone know if there is research being done about the huge similarities between this and psychedelics like LSD or Psilocybin mushrooms?
The images you get from said psychedelics can be extremely similar to these, so I guess we could learn something about our brain and how it recognizes and replicates patterns. Which part of the brain is responsible for this and how does it work, etc.
Mussi93 Looking at those images, the first thought that occurred was that they look surprisingly like stuff I've seen when taking psychedelics. Interpenetrating this from my own subjective experience, the similarities make me feel that this algorithm may be quite similar to part of the human minds function. Such a fascinating topic, I'd love to see more reporting on scientific research in this area, I can't get enough. The fractal patterns one sees under the influence of psychedelics must say something pretty important about how our mind works, just wish I could understand it more.
Mussi93 My intuition said's that the image you see on psychedelic drugs are the result of the pattern recognition systems in your brain going into overdrive, what should usually be filtered out or never recognized at all is being sent to the lateral areas of your brain instead. It would explain why things like faces are the most common pattern recognized while on psychedelics as that is the thing your brain put's the most effort into looking for.
@@SuperSilkyJohnson and eyes, i feel like that's a common theme, always fractal patterns with too many eyes. Amazing!
The part about the disembodied hands made me laugh.
Jesus that's freaking scary
This is so frickin' exciting. I'm loving some of the DeepDream creations.
This just blew my mind...
Plot twist: Google are actually developing an algorithm to solve those magic eye puzzles. :P
This is cool :) I think a way to improve would be to select the images from the image bank, say 100 of this type, 100 of another type, so there's not loads of dogs.
I've done some work in my career (and education) with neural networks. There's one error in this video: in general, you don't manually modify the nodes of a neural network. Instead you use one of many back-propagation techniques during the training process. The input to the network is an unknown image and the output of the network is "back driven" with the solution. This process modifies the weighting of the various nodes and their connections between and inside layers.
The engineer of the neural network might modify the number of layers, number of neurons in a layer, or even the types of neurons in a layer but they would never tamper with the nodes and their weighting algorithms directly.
I wrote a neural network in MatLab and trained it to recognize distorted letters (like in a captcha, but it was for old books with missing ink and water stained pages). The crux of it was that the training included human identification. So the neural network was as good as a person at identifying distorted characters, but no better.
This was an EXCELLENT explanation of what DeepDream is - I finally understand it! Thanks as always SciShow :)
This explanation is way cooler than the images by themselves. Neural networks rock.
The ad I had was Hank doing "I love science."I liked the video before the ad was over.
I love it when SciShow links to GitHub in their sources :)
Ladies and gentlemen, welcome to the dogscape.
This must be the coolest shit this channel has covered the last couple of years.
It is humbling to know that no matter how much you teach computers with all of their technological wonder, the human brain is still infinitely more powerful!
The images are nightmare fuel! lol
A step toward more sophisticated artificial intelligence, and computer generated art.
I can't look at the images. I always have a flashback to a bad acid trip immediately. it makes my stomach turn every time
These are the most accurate images of actually tripping I've ever seen
They should put those pictures in an art museum.
Crazy to watch this now. We went from this to Stable Diffusion in 7 years. It's crazy how this technology came out of nowhere. it wasn't even science fiction a decade ago, it was science fantasy. the idea of teaching an AI how to intuit and understand vague concepts? it's unbelievable.
I would love to see DeepDream work through the patterns of Hank's shirt!
Oh cool. I've seen some similar "artworks" of this before, didn't know they were called DeepDream.
I find that the phrase: "CATegories like DOG" blows my mind.
This is how computers see us. Give them pictures of hands holding mugs and you'll get a severed hand with your coffee.
I didn't want to watch the video now but I did because I was curious now I'll have nightmares thank you google
also(1) mildy disturbing. "mildly"(!!!)
also (2) I feel like this deepdream computer makes images like toddlers talk. you know when toddlers learn to talk and they say the words all messed up with like 10 extra letters... kinda like that.
This is all rather creepy to me. When I would try to fall asleep when I was a kid I would see images like those when I closed my eyes. I've since grown out of it, but seeing these things again makes my skin crawl. Thanks, Google!
It's ability to find the "dog" in a picture intrigues me, because it seems to open up possibilities to use computers to analyze and find patterns even more precisely than humans do. Just like humans could always draw curves, whole new possibilities opened up when we programmed computers to do that more perfectly. For instance we never be able to analyze the paths of particles in 3d space after a collision, we need to program computers to do that accurately. Similarly, there is probably something new we could get computers to do more accurately and precisely in the realm of previously holy human grounds of abstraction.
Ok so there are now computers that can somewhat accurately define pictures... well R.I.P to the various image based anti-bot securities.
Speaking of recognizing things, I was thinking about this recently: How do animals recognize objects? For example, let's say a human is walking outdoors with a walking stick, and an animal spots it in the distance. Does the animal recognize a bipedal animal with two arms that's holding a stick, or does the animal recognize a three-legged creature? Does it assume the stick is part of the human, or does it know that the stick is separate from the creature it's looking at?
'Creepy' is probably a good sign that we are getting close to something real.
Good timing!
That's interesting, training the nodes sounds a lot like how schema are modified in developmental psychology.
Even Stephan Hawking thinks "skynet" is a possibility. This is 1 step closer to it, this of a computer that can recognise what humans deem as safe or good hiding spots.
Oh no. The steps further into artificial intelligence are growing. What's that? SkyNet? I can hear you off somewhere in the distance!
Thanks Hank. Now I will have nightmares... forever!
DeepDream is such an interesting piece of computing, I wonder what this collected data will lead to in the long term...
Joseph Stalin AI overlords.
the other night i was reading a wikipedia page about Operation Fishbowl. Could be an interesting topic for an episode some time.
Looking at deepdream pictures make me feel anxious.
One step closer to the robot takeover
Just what i was thinking
Alex Bracau Wrong, its a step into the world of computers tripping balls.
CorruptHumanoid i literally just pictured a bunch of Mac books and Dell laptops chilling together passing a bong around.
KaylaLoveYaLucas More like a conference room full of different computers eating shrooms that grow on the ceiling.
I remember when they had a game where you had to classify the images. So basically I think they used people to do their dirty work early on.
So now computers can make surrealist art, cool.
I am getting so excited! Sentient computers may come to exist in my lifetime!!!
What about sentient Microwaves? Constantly judging you on your eating habits and asking if you know any single microwaves wanting to link GUI's for a computer night.
2045-2090 according to the experts.
The later they arrive, the more work we will have put into friendly AI though, and the less likely it is to use our atoms for something else.
Though if you tell more people about Roko's Basalisk, those times should drop a bit.
Sergio Garza Do you have one of these? Because I have a toaster that needs to let off some steam.
As a Microwavable food product I find it intriguing to "get inside" any single microwaves in my area.
MicrowavableToast
"Very Compatible Units In Your Area!"
I read it as "Derp" dream
This reminds me of when I was little and when I would stare at the patterns on a tile floor looking for the group of splotches that looked the most like a face.
3:42 why did this make me cry of laughter
*have. (grammar correction scishow). :) love u guys
I'm pretty sure this is how Skynet became self aware.
My brain does not compute how this works but it seems cool.
You know, I once did a deep dream search of a colossal titan. I wish you could see it.
I remember reading positive stories about neural networks last century. Seems it still has not come very far since then.
Really interesting stuff. I'm interested in how we're actually working in creating smarter AI. Feel free to make more videos about this topic, and maybe some references to movies like Ex Machina or even Chappie.
what can be seen cannot be unseen
Love your channel! Any chance you guys are gonna talk about this 'self aware' robot? I'd love to hear your insight if you do. Thank you!
Like IceMetalPunk said, it's quite possibly just like imagination is side effect of human pattern recognition abilities.
But it's also quite profound. Even if this doesnt lead to A.I directly (although it could), the fact that researchers already need to probe their system to understand how it works means that, once A.I emerges, people will have a hard time realizing that it had.
Vaidas Šukauskas A lot of technologies were invented, but took many years before they were fully integrated into society. With how the Internet is approaching a sort of "cultural singularity", and how increasingly interconnected we are, I wouldn't be surprised if AI *is* developed but rather than changing everything all at once, it gradually seeps into our lives. The philosophical questions are asked one at a time.
That sort of situation, where you don't know how a piece of software works until you probe it, is somewhat unique to all learning, machine or otherwise. When a human learns, synapses form and rearrange in the brain, and we can't yet look at those structures and immediately infer how they work (though we continually make steps toward that goal). We need to ask them what they're doing--"probe" them. The same is true here: the neural net is constantly adjusting its weights, but to humans, those are just a bunch of numbers. It's extremely difficult to look at the organization of numbers and directly infer how the network is running. So instead, we "ask" it to walk us through its process with an output at each step, the same way we'd ask a child how they worked out their homework exercises.
When general AI is born, we will know immediately, because it's defined not by its methodology, but by its behavior. And the behavior of a system is far more visible to us than the processes that lead to said behavior.
What I find most disturbing about google's daydreaming software, is many of them make me extremely uncomfortable with how close they are to a couple... college experiences. I didn't even know it was possible to portray that in an image...
+Richard Smith ever hear of Alex Grey ? great artist , definitely has done ayahuasca ceremonies in which one travels with the use of DMT from plants in the Amazon rain forests .. it is a naturally produced chemical stored in the brain and released at birth and death ( as well as near death ie: stories of light at the end of a tunnel ) .. the long and short is that the images produced by deep dream are quite interestingly more than similar to what one sees during such a journey
I really hope they come out with different publically-accessible variants of Deep Dream designed specifically for detecting and enhancing different patterns than just dogs and eyes for such things to produce different aesthetics from the photographic enhancement.
Like, an industrial one designed for enhancing machines & tools and smoke in things, a horror one designed specifically to enhance blood; gore; skulls & monsters, one perverse one designed for enhancing furries; tentacles & genitalia because god is officially dead and I am trash. Would be neat from an aesthetic standpoint is all i'm saying.
I wish I found your channel ages ago
Well that explains the unusual obsession with dogs.
Cool and frightening technology! The name 'Deep Dream' sounds like an X-Files episode.
Look how far we have come
hank went to the barber and was like "just fuck my shit up"
Oh dear, we're teaching computers to imagine now, this really is the beginning of the end.
1:51 Manually doing what the brain does automatically? I think we may be on the right track.
Hank.. your shirt looks like one of those pictures =P
Getting alot closer to creating a terminator.
With all the information about everyday objects these servers have, they should put them up against Re-Captcha!
This is absolutely amazing!
The comps are trolling us with all that animal imagery
*What everyone sees:* A robot imagining something artistic and creative.
*What it really is:* Random noise produced by a confused machine that has no clue what it's even doing.
Man we've come so far so fast
Shit looks like some acid trips I had back in the day
This gives me the absolute creeps.
Thes images resembles those Hypnagogic Halucinations; when you first open your eyes after waking up, and you see objects transfigurated into something else for one or two seconds.
SciShow, could you do an episode on Planned parenthood selling baby body parts and how it is a wrong way to get stem cells?
My own dreams resemble these more than it does inception... But at least my dreams make aesthetic sense. These images don't, but they are very enjoyable in a vicarious drug trip sort of way.
Will you go over the science of the Sony project N prototype (parabolic speakers emitting sounds to your meatuses while not needing ear phones in your...meatuses)????
Also, something to do with voice recognition, like how a computer converts analogue voice or sound into binary through the use of sampling, and compression/decompression algorithms???
Maybe how Microsoft Kinect works???
All sound good.
Ha, I gave a big post on one of scishow's videos explaining neural networks before. More in the context of game-playing AI, though, rather than image recognition.
MILDLY DISTURBING?
MILDLY
Vapourwave now has images. It's all becoming a reality.
im convinced that deep dream makes everything a dog
Look, scientists, it's really cool that you're trying to make computers recognize faces, but this is a drug trip simulator
Nice demonstration how to _project_ meaning where there is none. At this rate, we might have the first robot religion on our hands in the next 20 years.
I wonder if AIs of the future will look back on these experiments with... disapproval.