How Humans Affect Machine Learning: Privacy, Efficiency, and Biases

Abstract

Artificial intelligence increasingly intersects with human lives, both by processing personal data and by influencing societal systems, such as decision-making. These interactions necessitate careful considerations of privacy preservation and the mitigation of societal impacts, such as algorithmic biases in decision-making. Achieving these objectives requires leveraging both computational and human resources effectively during the decision-making process. This thesis addresses these challenges, focusing on the private generation of data that preserves statistical characteristics while maintaining privacy, as well as optimizing models that integrate human input in decision-making processes.

In the realm of privacy preservation, this work introduces a novel approach to summarize and privatize data distributions while enhances the quality of the generated data by compressing the distribution in an embedding space. For decision-making systems involving human resources, referred to in this thesis as learn-to-defer (L2D) methods, this thesis introduces methods to train these models in an active and offline manner and analyzes and compares their sample complexity. This work further achieves a uniquely optimal solution for L2D systems with secondary objectives such as algorithmic fairness. Finally, it extends the idea of L2D to systems where the human expert and model prediction can be combined. Through these contributions, this thesis advances the state of the art in privacy-preserving data generation and human-AI collaboration, addressing technical and societal challenges in the deployment of machine learning systems.