Episode 2
The weather forecaster
Why should a model's confidence mean something?
Listen · 10 min
Show notes
A seventy percent chance of rain from a weather application is reliable because it has been checked against the world. Vinay then asks why we should not demand that confidence scores on AI be checked the same way. He walks through the 2017 paper that found that deep learning models were overconfident by default, the single number fix that mostly cured that overconfidence, and how he draws a calibration curve for his own models. He closes with the diagram that he would want a regulator to require in a public description of a device.
Sources mentioned
- Guo, Pleiss, Sun, and Weinberger, "On Calibration of Modern Neural Networks," Proceedings of Machine Learning Research, ICML, 2017, https://proceedings.mlr.press/v70/guo17a/guo17a.pdf
- Guo, Pleiss, Sun, and Weinberger, "On Calibration of Modern Neural Networks," arXiv preprint 1706.04599, 2017, https://arxiv.org/abs/1706.04599