EECS Seminar: Human Decisions and Machine Predictions
Speaker
Sendhil Mullainathan
Chicago Booth
Host
Aleksander Madry
Abstract:
Machine learning is increasingly being used to influence human decision-making, even in high-stakes decisions such as those made by doctors and judges. Early results are decidedly mixed, showing both high risk and high reward. On the one hand, deployed systems have failed, sometimes at immense scales and with drastic consequences. On the other hand, empirical results suggest that - as humans too are biased - well-designed algorithms can have large positive social impact.
In this talk I will illustrate in some more detail the risk and reward of these systems, but also propose ways to improve our existing frameworks for building and evaluating them. The latter will involve two specific approaches: use of econometric causal inference tools to improve how we evaluate algorithms; and use of behavioral science insights to change how we build them. These perspectives on people and data turn out to also inform a broader set of uses of machine learning, such as building tools for self-insight or scientific discovery.
Machine learning is increasingly being used to influence human decision-making, even in high-stakes decisions such as those made by doctors and judges. Early results are decidedly mixed, showing both high risk and high reward. On the one hand, deployed systems have failed, sometimes at immense scales and with drastic consequences. On the other hand, empirical results suggest that - as humans too are biased - well-designed algorithms can have large positive social impact.
In this talk I will illustrate in some more detail the risk and reward of these systems, but also propose ways to improve our existing frameworks for building and evaluating them. The latter will involve two specific approaches: use of econometric causal inference tools to improve how we evaluate algorithms; and use of behavioral science insights to change how we build them. These perspectives on people and data turn out to also inform a broader set of uses of machine learning, such as building tools for self-insight or scientific discovery.