|
Hinton's deep learning and other approaches have given rise to really good generative models for natural images. I am looking to get the learned parameters and the model in a convenient format so that I can use it for, say classification or some other goal without having to burn CPU time learning the parameters. I am just looking for a model that is either relatively easy to implement (at least the generative forward version or probability evaluation, not the leaning method), and pre-learned parameters that are rich. It doesn't have to be Hinton's work. Does anyone know if these are available? |
|
Generating realistic images is very difficult. I only know of ways to generate images from a highly constrained class, or very small patches of images. The best probabilistic model of large natural images I know produces samples that like like abstract art at best. Le Roux et al.'s Learning a generative model of images by factoring appearance and shape is an interesting approach, I think, but the samples they show at the end don't look very 'natural' either.
(Mar 16 '12 at 20:18)
Sander Dieleman
@Sander Thanks for the link. Didn't know that one :)
(Mar 16 '12 at 20:27)
Andreas Mueller
|
You can check out fields of experts. Though I'm not sure if they are any good for classification. Usually they are for impanting. Why do you want a generative model for classification?
Thanks! While I did not mention it, I am actually more interested in actually generating images. I see that http://www.gris.informatik.tu-darmstadt.de/~sroth/research/foe/downloads.html has trained models. Do you think they can be used for generating images?
As far as I know, it is not possible to generate realistic images at the moment. Having interesting "patches" is the most you can ask for. Hinton's group can generate MNIST digits and NORB-like images but natural images are another story. For Fields of Experts and related model, I think the idea is more that you have some parts of an image and want to fill in missing parts. Examples are noise removal and "inpainting", where you fill in parts that where overwritten, for example by some text on an image.
NORB image * random scaling * random colors + layers of NORB images = sort of natural image? Do you know if Hinton's group has downloadable learned parameters to generate NORB images.