Student enrolled in his Bayesian Machine Learning class ORIE 6741. The formulation of adding bayesian inference to deep learning can be applied to CNN's as well. I plan on using his work on Stochastic Gradient Hamiltonian Monte Carlo to do Bayesian Inference on Image Segmentation as well. For more interest, check out this blog, super interesting: http://www.cs.ox.ac.uk/people/yarin.gal/website/blog_3d801aa...

I'd like to flag up this paper on HMC and minibatch inference: https://arxiv.org/abs/1502.01510v1. It seems that SGHMC might not always behave as well as you might like.

I should say this isn't to pour cold water on the idea of Bayesian deep learning - I think it's an extremely sensible idea and undoubtedly a good direction for research, but just wanted to caution that this particular MCMC method might have subtle flaws.

I wished more academic papers would have a "Limitations" section. Often it is so hard to tell what these are, e.g.: Does it scale well? How much faster/slower is it than other methods? Does it converge reliably?

No incentive for such a thing.

Neither is there for reproducible code, but now there are artificial incentives such as the Reproduction Challenge.

Did you read "Variational Gaussian Dropout is not Bayesian", https://arxiv.org/abs/1711.02989?

https://twitter.com/dustinvtran/status/928493076854669312

I posted before refreshing and missed the link to this paper. Will need to read through it.

Interesting blog post! From a 10,000 ft overview is this method utilizing the conditional posterior's to incorporate uncertainty information into the model GAN model? If so how does it differ from using drop-out's to approximate Bayesian behavior ala https://arxiv.org/pdf/1506.02142.pdf ?

For a super quick (non-expert) explanation of a Generative Adversarial Network, you essentially have 2 models that are pitted against each other:

1. A generator ANN creates synthetic examples 2. A evaluator ANN indicates how "realistic" the examples are

The generator uses the evaluator output to incrementally improve it's synthesis.

And for Bayesian GAN, you have a model that generates those two models that are pitted against each other.

What does it do with the models? So it samples a probability distribution, by training GANs, but how does it evaluate the points? (Which GAN is better?)

Even after looking at the readme, I don't understand how it works. Could you explain it a bit, please?

If you read the paper https://arxiv.org/abs/1705.09558, in section 2.1, it defines two conditional posteriors for the parameters of the generator and discriminator. Then, classical GAN is just a maximum likelihood estimate (or, a MAP estimate with uniform prior) of the parameters.

Next, the paper said: how about sampling the whole posterior distribution instead of finding only one point of maximum likelihood as classical GANs do. This is the point where the famous Markov chain Monte Carlo (MCMC) algorithms become useful. They use something called stochastic gradient Hamiltonian Monte Carlo, basically, it is a random walk algorithm, at each step, you follow a noisy gradient, as a result, you converge to the posterior distribution instead of a local minima as gradient descent does.

The paper claims that sampling the whole posterior helps to resolve problems with classical GANs. IMHO, this isn't a surprise claim, this is exactly what is good about Bayesian statistics.

Is it fair to presume that we'll next see a variational inference approach to this, with faster but slightly less optimal results?

I believe it models a distribution of the weights the network after training. Once you can generate samples, you can average the values to get expected mean, which is the improved prediction.

And the evaluator ("discriminator") uses the generated outputs to improve its discrimination.

Very cool! Can't wait to try this with BEGAN! I am curious if this could model some probabilistic decision making as well outside usual image/video application, like with complicated/unknown (inverse) reinforcement learning policies.