Feature rescaling needed for bayesian neural networks?

[ferrine]

I think absolute values may be hard to extrapolate and first differences dor sequential data might be a good feature regardless Bayesian stuff.