Look into adding soft KNN for monitoring during training

[owenvallis]

https://arxiv.org/abs/1902.01889

[owenvallis]

"As an analytical tool, it provides insights into the evolution of class similarity structures during learning"

[igor-susic]

Hi @owenvallis if this one is still for grabs I would take some time to implement it.

[innat]

ref. https://github.com/gaudelbijay/SNNL-Loss/blob/main/soft_nearest_neighbor_loss.py