Bayesian neural network

Dear EI-Team,

I am working on a regression problem. One of the questions that I like to answer is how “accurate” is the predicted value. One of the ideas is to implement a probabilistic Bayesian neural network to obtain an idea about the mean and variance, but also a confidence interval (CI) of the prediction, and if this approach can be used to detect Out-of-Distribution (OOD) examples.

My question: Is it possible to deploy this model on a device?

Any suggestions on other approaches/ideas are always welcome.


Hi @Joeri,

We haven’t yet experimented with converting Bayesian networks via our tooling. I had a quick play around with this model to investigate:

Superficially it looks like we should support all the operators it uses, but the best way will be to give it a try with BYOM :slight_smile:

Will be interesting to see what you find!


This is what I need to know.

I am also following the Keras reference you mention. I am working on a baseline standard neural network (using Edge Impulse Python SDK, together with wandb); once this is fine, I try to translate it into a BNN. I will keep you posted.

Thanks for the feedback.


Very cool, good luck and let me know!