How modeling choices affect machine learning predictions

All machine learning models make assumptions in order to learn statistical regularities from data. In order to trust the results of the machine learning analysis, it’s critical to assess how much the results depend on the initial assumptions made in the modeling process. Often, users are faced with assumptions which are qualitatively interchangeable: there is no reason to prefer one assumption over another, given prior beliefs. In these cases, large discrepancy in the results due to switching between assumptions is a cause for concern, particularly when decisions are made based on those results. In this talk, Dr. Ghosh will focus on Gaussian processes, a class of models widely used to analyze spatial and time-series data. He will explore the sensitivity of their predictions to modeling choices and illustrate how innocuous changes in assumptions can have a large effect on the model’s predictions, using a NOAA dataset of CO2 emissions. Learn more about IBM Research AI: