Bayesian Neural Net

[Pietro Portolani]

Hi everyone,

I've implemented in Edward2 the example of the Bayesian Neural Network found in the Edward tutorials (http://edwardlib.org/getting-started). Now i'm testing it on the toy data of the Gaussian Process tutorial (https://github.com/tensorflow/probability/blob/master/tensorflow_probability/examples/jupyter_notebooks/Gaussian_Process_Regression_In_TFP.ipynb).

I was expecting a behaviour similar to the Gaussian Process one: outside the definition range of the data, the prevision as an high variance. This doesn't happen, Was I expecting wrong? Any suggestions to get that results?

Thanks in advance,

Pietro

[Dustin Tran]

Thanks for reaching out! For various reasons, I left the team and actually don't work on TensorFlow Probability anymore (including the Edward2 as-is in that repository). So I may not be the best person to help diagnose the user question.

With that said, one helpful reference could be Danijar's code (https://github.com/brain-research/ncp). It uses an extension of TensorFlow Keras layers to implement variational BNNs, and the toy figures do indeed show high variance outside the distribution of the data inputs. But it depends on the model—variational BNNs don't guarantee it. Software-related, one of our plans is to redesign the layers to something we can iterate a bit more easily with, both in that code base and in other Bayesian neural net work. Hopefully that might also be useful for whatever you're using BNNs for!

Dustin

--
You received this message because you are subscribed to the Google Groups "TensorFlow Probability" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfprobabilit...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tfprobability/.

[Pietro Portolani]

Thank you very much, I'll definitely have a look! Good luck for your next project!

Pietro