New Chapter: Uncertainty

[whitead]

Bayesian, Deep ensembling, Calibration, dropout, flows?

[kjappelbaum]

• Conformal predictions (e.g., https://arxiv.org/abs/1908.03569) are also quite interesting
• Gives a nice bridge to the infinite width limit of NN and stuff like NNGP and NTK
• The cheminformaticians had some "domain of applicability measures" that are somewhat related and often distance based (and in this way similar to the latent space based approach from the Kulik group)
• would be sweet to mention epistemtic aleoteric uncertainty
• might also be interesting to mention uncertainty calibration (Platt scaling, isotonic regression)

[whitead]

https://twitter.com/balajiln/status/1421519413988053002?s=21

[whitead]

https://twitter.com/mlstjournal/status/1429017793119281158?s=21

[ziyueyang37]

comment

[kjappelbaum]

https://arxiv.org/pdf/2110.04629.pdf and the library