These models are just clever image uncompression. Nothing scary here. :)
It's just like using Markov chains to make English-like gibberish, but for pictures.
The interesting bit is that (I think) the generative adversarial network is regressing from random noise to an image, which isn't how most autoencoders work.
> These models are just clever image uncompression. Nothing scary here. :)
Yes and no. The latent representation space can be seen as a kind of result of compression, but the opposite is going on. As you say, these models generate images from (latent space) random noise, but there's no direct mapping from image space to latent space (as there are with autoencoders), so no way to compress, really.
If you insist on a practical way to do it, gradient descent on the latent vector would probably work well given the type of results DeepDream has obtained.
Compression is a subtle and powerful thing. The ability to compress is closely related to prediction: if you can predict 95% of the moves a chess grandmaster will make, you can compress the game by explicitly representing only the other 5% of moves. If you could perfectly predict (or equivalently, compress) the actions of a real-world human being, you'd have solved AI.
Despite their simplicity, Markov chains are used throughout modern statistical AI, e.g., the Google Translate language model is essentially a big Markov chain. The fact that deep networks can apparently form better generative models across a wide range of applications (no one has ever actually gotten these kinds of image generation results from Markov chains) means that they really are getting at more interesting structure. They're not a panacea, but it's still a pretty big deal.
The Latent space of the image vectors is rich with emergent semantic algebra.
Extracting a vector for glasses wearing and adding it to another person; rotating a face is a consistent vector; a vector for removing windows from any imagined room; a vector for smiling.
This all seems a bit like Geoff Hinton's thought vectors - nice work.
Looks very intersting (and beautiful), wish I could understand some more of it :)
Could someone explain in simple terms how the arithmetic work? E.g. when they do "smiling woman" - "neutral woman" + "neutral man" = "smiling man", is it like you perform the operations on the networks' weights or something?
I think AI-generated art is an industry waiting to happen. Imagine telling a VR system "show me a mountain ski resort with skiers going everywhere" and it just generates it. It could have a huge impact on procedurally generated maps in gaming.
Or even plays or fiction! Look at how formulaic detective shows are. It'll probably be hard to get the emotional nuance, but maybe AI could make all of the characters and plots.
While I think we're far away from generating TV shows with Conv-Nets, perhaps it can do simpler things (porn?).
>It could have a huge impact on procedurally generated maps in gaming.
I think it could go far beyond just maps. A major immersion breaker in most open-world games is the lack of detail in the lives of non-player characters. It's impractical to manually create this content at any real scale, so I think that's an area where AI would shine.
However, that isn't terribly valuable without also being able to generate convincing conversation dynamically—and that's definitely Turing Test territory. It would also help to synthesize that output in the form of an uncanny and unique human voice, which is also hard, and probably squarely in the same territory.
That said, I think all of this could be accomplished relatively soon, and without general intelligence necessarily being solved.
Of course, once general intelligence is solved (and doesn't go sideways on us), it's a good bet that this kind of thing will be nothing short of incredible. I imagine the Holodeck from Star Trek might be a fair approximation.
As an aside, utilizing general intelligence within interactive entertainment mediums for the purpose of creating believable characters may prove to be highly unethical. Bostrom's Superintelligence touched on this, and it's quite interesting if not terrifying to ponder.
It would be interesting to see how you can do machine learning approaches for the kinds of descriptions WordsEye handles using rules, like "the man is in front of the wall." How do you learn to generate an image from that kind of description? You might need to have some kind of learning in abstract 3d space, you might not.
27 comments
[ 0.73 ms ] story [ 76.5 ms ] threadIt's just like using Markov chains to make English-like gibberish, but for pictures.
The interesting bit is that (I think) the generative adversarial network is regressing from random noise to an image, which isn't how most autoencoders work.
Yes and no. The latent representation space can be seen as a kind of result of compression, but the opposite is going on. As you say, these models generate images from (latent space) random noise, but there's no direct mapping from image space to latent space (as there are with autoencoders), so no way to compress, really.
If you insist on a practical way to do it, gradient descent on the latent vector would probably work well given the type of results DeepDream has obtained.
Despite their simplicity, Markov chains are used throughout modern statistical AI, e.g., the Google Translate language model is essentially a big Markov chain. The fact that deep networks can apparently form better generative models across a wide range of applications (no one has ever actually gotten these kinds of image generation results from Markov chains) means that they really are getting at more interesting structure. They're not a panacea, but it's still a pretty big deal.
http://googleresearch.blogspot.ca/2015/08/the-neural-network...
There is lots of information about the 3D world present in the light (at least for the bedrooms).
This could be very useful for spatial navigation or computer vision, I'd think.
Paper: http://arxiv.org/pdf/1511.02570.pdf
He's in that research group, anyway.
Extracting a vector for glasses wearing and adding it to another person; rotating a face is a consistent vector; a vector for removing windows from any imagined room; a vector for smiling.
This all seems a bit like Geoff Hinton's thought vectors - nice work.
http://cs231n.stanford.edu/reports.html
This is the moment I am reminded of how much down the foodchain I really am.
Could someone explain in simple terms how the arithmetic work? E.g. when they do "smiling woman" - "neutral woman" + "neutral man" = "smiling man", is it like you perform the operations on the networks' weights or something?
We are doing vector arithmetic in this input space, and the outputs seem to be semantically changing according to the vector arithmetic.
Buh.
Or even plays or fiction! Look at how formulaic detective shows are. It'll probably be hard to get the emotional nuance, but maybe AI could make all of the characters and plots.
While I think we're far away from generating TV shows with Conv-Nets, perhaps it can do simpler things (porn?).
I think it could go far beyond just maps. A major immersion breaker in most open-world games is the lack of detail in the lives of non-player characters. It's impractical to manually create this content at any real scale, so I think that's an area where AI would shine.
However, that isn't terribly valuable without also being able to generate convincing conversation dynamically—and that's definitely Turing Test territory. It would also help to synthesize that output in the form of an uncanny and unique human voice, which is also hard, and probably squarely in the same territory.
That said, I think all of this could be accomplished relatively soon, and without general intelligence necessarily being solved.
Of course, once general intelligence is solved (and doesn't go sideways on us), it's a good bet that this kind of thing will be nothing short of incredible. I imagine the Holodeck from Star Trek might be a fair approximation.
As an aside, utilizing general intelligence within interactive entertainment mediums for the purpose of creating believable characters may prove to be highly unethical. Bostrom's Superintelligence touched on this, and it's quite interesting if not terrifying to ponder.
It has been happening for a while:
http://www.pamelamc.com/html/aaron_s_code.html https://www.wordseye.com/
It would be interesting to see how you can do machine learning approaches for the kinds of descriptions WordsEye handles using rules, like "the man is in front of the wall." How do you learn to generate an image from that kind of description? You might need to have some kind of learning in abstract 3d space, you might not.