Bias Generation and Amplification
Developing a theoretical framework to identify the key factors contributing to ML misbehaviour against population subgroups.
Description
This project investigates the theoretical underpinnings of algorithmic bias. By leveraging tools from statistical physics and high-dimensional probability, we aim to understand how data geometry and training dynamics contribute to the generation and amplification of bias in deep learning systems. This work is supported by OpenAI’s Alignment Team via the AI Alignment Project.