I am testing bayesian neural networks to see how they perform when compared with neural networks created with keras.
Some of the features are temperatures which increase and descend and others are cumulative counts of certain events can keep increasing.
1. Is it needed to re-scale the features between 0 and 1 during training?
2. What if the maximum values of the temperature and cumulative features are significantly higher than during training?
Hi, what kind of BNN do you deal with?
What do you mean with what kind?
The Keras NN I have made is a CNN with 4 conv1d layers.
It is used for classification.
I think absolute values may be hard to extrapolate and first differences dor sequential data might be a good feature regardless Bayesian stuff.
BTW, does non-bayesian net work for you?
Although I will try re-scaling too.
But what I am worried about is e.g. when in my training set one of temperature features may be 80 degrees at max but during the prediction it may be 120 degrees.
Would retraining be needed in that case?
I did not quite understand this:
Could you explain a bit?