regressing multimodal distributions accurately

[jpata]

Something Michael K said in the meeting about using simple regression for multimodal target distributions being a poor approximation and only able to predict the avergae.

I believe this could have a big effect, as we have many different particle types with different kinematic spectra (e.g. energy), yet the regression network is the same and just tries to predict a single value.

Need to find a good reference for this.

[jmduarte]

https://towardsdatascience.com/anchors-and-multi-bin-loss-for-multi-modal-target-regression-647ea1974617

Multi-bin loss is interesting

[jpata]

Despite my best efforts in #234, in the end I have reverted the multi-bin / multimodal loss to simple regression in TF #253, because despite promising initial checks, the final physics performance after a long training was not competitive.