Well this is delightfully terrifying. Neural Networks that can create environments and people on their own.
back
2 comments
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.
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.
Pretty sure Google Translate is not a Markov model at this point; pretty sure it's a deep recurrent network.
> 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.
Of course there is: direct enumeration.
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.
I think variational autoencoders (which explicitly model the compression step and the decompression step) seem like the more classical way to do it.
Ya... Also they said that the code will be available today on the same repo (on the master branch, the current link is on another branch).