SVI trainer for Bayesian neural networks #273
Description
Is your feature request related to a problem? Please describe.
The previous Bayes-by-Backprop (BBB) trainer in UQpy relied on a fixed or externally specified noise term, which limited its ability to represent learned observation noise. This restriction made it difficult to properly quantify inherent noise when training Bayesian neural networks.
Describe the solution you'd like
Extend the existing BBB trainer to SVI trainer to include a trainable inherent noise parameter, allowing the model to jointly learn both weight distributions and the observation noise scale during training.
This approach enables the network to capture uncertainty directly from data .
Describe alternatives you've considered
Extending the existing BBB trainer to SVI.
Additional context
Add any other context or screenshots about the feature request here.