Computational Insights on the Meaning of Individual Probabilities

Prediction algorithms assign numbers to individuals that are popularly understood as individual “probabilities”—what is the probability of 5-year survival after cancer diagnosis?—and which increasingly form the basis for life-altering decisions. The philosophical and practical understanding of individual probabilities in the context of events that are non-repeatable has been the focus of intense study for decades by the statistics community. The wide-scale impact of automatic decision making calls for revisiting these questions from a computational perspective.

In this vein and building off of notions developed in complexity theory and cryptography, we introduce and study Outcome Indistinguishability. Predictors that are Outcome Indistinguishable yield a model of probabilities that cannot be efficiently refuted on the basis of the real-life observations produced by Nature.

The talk will be self-contained and will explain the relevant complexity-theoretic and algorithmic-fairness literatures in which this work is grounded. Our focus will be on the insights that can be drawn from this work such as providing scientific grounds for the political argument that, when inspecting algorithmic risk prediction instruments, auditors should be granted oracle access to the algorithm, not simply historical predictions.

Based on research joint with Cynthia Dwork, Michael Kim, Guy Rothblum and Gal Yona.