Validity and Efficiency of Set Predictors
Standard algorithms of statistical machine learning produce predictions that are not accompanied by any measures of their reliability. Such measures are provided in traditional statistics, but under different (often much more restrictive) assumptions than the assumption of exchangeability (the data are produced independently from the same distribution), that is standard in statistical machine learning.
This and the following talks are devoted to the method of conformal prediction, which has been designed to bridge this gap. Conformal predictions incorporate measures of their own reliability and are automatically valid under the exchangeability assumption.
In this talk I will concentrate on the remaining desiderata for conformal prediction, namely, conditional validity and efficiency.