- Tags:: Book (still not read)
- Author:: Aston Zhang, Zach C. Lipton, Mu Li, Alex J. Smola (Amazon Data Scientists)
- Link:: Dive into Deep Learning — Dive into Deep Learning 0.17.2 documentation (d2l.ai)
- Source date:: 2021-01-01
- Finished date:: 2021-01-06
4.9 Environment and Distribution Drift
Covariate shift: That p(x|y) remains constant is an assumption we make, and we do so in circumstances where we believe that x causes y. Label shift: That p(y|x) remains constant is an assumption we make in cases where we believe that y causes x (disease → symptoms). Model drifting: We may also have a third case, Concept drift, where the very definition of “y” may change. The only thing we can do is exploit knowledge of those variations being very gradual in time, space…