Charla "Evaluating ML: When We will Stop Fooling Ourselves?"

Expositor: Ricardo Baeza-Yates (Universidad de Chile)

Resumen: ML Evaluation is usually based on an average measure of success such as accuracy. This kind of evaluation has several drawbacks: (1) the model works well for easy instances but badly for difficult ones, but the actual real distribution is usually not known; (2) this assumes that all errors have the same impact, which is almost never true; and (3) optimizing success does not minimize critical errors. In this presentation we discuss these problems and give some solutions that address them.

--
Comunicaciones DCC