Proposal Pre-Writing by Alyssa Stephens https://padlet.com/adsteph/shsuopjz1lci5rf7 Made with a stroke of good luck en-us 2020-10-26 14:14:34 UTC 2020-11-07 20:50:30 UTC hello@padlet.com Interpretability methods allow ML to be used to learn new science, rather than just a prediction. adsteph https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862523220 2020-10-26 16:33:53 UTC https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862523220 Some states of the system (or input samples) are more predictable than others adsteph https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862527454 2020-10-26 16:34:40 UTC https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862527454 Knowledge → ML adsteph https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862662732 1) Include “K” in choice of inputs.  Ex.: Knowledge-guided selection of input variables, pre-processing, but also topological data analysis.
2) Include “K” in loss function of NN.  Ex.: Add loss term for physical constraint.
3) Encode “K” in model selection.  Ex.: Physical constraints enforced in NN architecture. 
]]> 2020-10-26 17:01:22 UTC https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862662732 Backward direction: ML → Knowledge adsteph https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862666193 4) New “K” from studying how a trained model maps input → prediction/graph  (does not require interpretability).  Ex.: synthetic inputs.  Graph neural networks.  Causal discovery.
5) New “K” from studying where a trained model is focusing in input when generating prediction.  Ex.: attribution heatmaps.
6) New “K” from studying individual components of trained ML model.  Ex.:  interpretation of latent space???]]> 2020-10-26 17:02:02 UTC https://padlet.com/adsteph/shsuopjz1lci5rf7/wish/862666193