The Need for Interpretability Biases

In his seminal paper, Mitchell has defined bias as "any basis for choosing one generalization over another, other than strict consistency with the observed training instances", such as the choice of the hypothesis language or any form of preference relation between its elements. The most commonly used form is a simplicity bias, which prefers simpler hypotheses over more complex ones, even in cases when the latter provide a better fit to the data. Such a bias not only helps to avoid overfitting, but is also commonly considered to foster interpretability. In this talk, we will question this assumption, in particular with respect to commonly used rule learning heuristics that aim at learning rules that are as simple as possible. We will, in contrary, argue that in many cases, short rules are not desirable from the point of view of interpretability, and present some evidence from crowdsourcing experiments that support this hypothesis. To understand interpretability, we must relate machine learning biases to cognitive biases, which let humans prefer certain explanations over others, even in cases when such a preference cannot be rationally justified. Only then can we develop suitable interpretability biases for machine learning.