Negative loss while using BNI implementation option

[CeceliaSoh]

Hi @jiawei-ren
I am trying to implement the Balanced MSE to other long tail regression task. Firstly, I tried the BNI, and then the negative traning loss are often computed and they are used to update the model. I would like to ask the appearance of negative loss is usual while using the BNI?

[jiawei-ren]

Hi @CeceliaSoh, sorry for the very late reply. Yes, it is normal.

BNI is equivalent to a negative log-likelihood, which should be positive. However, I omitted the constant bucket weight in the numerator (i.e., the MSE term) so that it could go negative in my implementation.