Restricted Boltzmann Machines vs GAN

Question

Can someone please tell me how RBMs and GANs compare to each other? I know that the community is more excited over GANs than RBMs at the moment. I guess it's because GANs produce better results? My question is this

• Can RBMs be a replacement for GAN in the sense, for example, can it generate images similar to the photo-realistic(kind of) images created GANs?
• Are there any advantages to using RBMs over GANs? Maybe less training/generation time? Less memory footprint?
• How does RBMs compare to GANs in general?

Thanks!

Answer 1

How does RBMs compare to GANs in general?

RBM is sort of a layer whereas GAN is an architecture. GANs are composed of two parts for which you can use different networks - convolutional, fully connected et c, whereas RBMs define joint probability on one layer of visible and hidden units.

Can RBMs be a replacement for GAN in the sense, for example, can it generate images similar to the photo-realistic(kind of) images created GANs?

That would be really hard given that GANs that do this thing use convolutional neural networks, which are more suited for image data than RBMs.

Are there any advantages to using RBMs over GANs? Maybe less training/generation time? Less memory footprint?

GANs are known to be hard to train. That said, they are more general (since you can plug in anything you like as encoder and decoder) so this comparison doesn't make much sense (it's sort of like comparing linear regression to neural networks for regression).