behavior of the epistemic uncertainty from the loss is weird for the toy-dataset

[AdithyaVenkateshMohan]

https://colab.research.google.com/drive/1iRu82bcKd1aF8DgZB_x-deIJbt-GNf-R?usp=sharing


as u can see the above image uncertainty from the positive side of the graph seems to be reducing even though it should be increasing similar to left side of the graph

[AdithyaVenkateshMohan]

@aamini

[kpqn]

Has your issue been resolved? I have a similar question as well.

[ZZY18]

The cause of this problem is speculated to be due to the activation function that selects the Relu.(I observed a difference in the negative values of your data). At the same time, I also found this asymmetry in the prediction case given, which can be alleviated by changing the activation function to Tanh.


I hope I can give you some help！