Measuring and Mitigating Biases in Vision and Language Models
Abstract:
Remarkable progress has been achieved in many vision and language tasks, however, recent studies have revealed that many proposed models exhibit various biases. For example, a human activity recognition model can overly correlate man with coaching and woman with shopping. Great concerns have been raised about the potential adverse effect of such correlations on societal fairness and equality. As more and more research techniques are being adopted in practical applications, it is critical for us to be aware of how biases exist in datasets and models and how to mitigate them. In this proposal, we approach the problem by 1) establishing metrics to quantify biases; 2) constructing benchmark datasets to evaluate biases; 3) introducing specific approaches to alleviate biases. The completed work demonstrates several methods we proposed to measure and mitigate gender bias in object recognition and word embedding systems. Building upon the early results, we will next focus on more general biases. We seek to build more fair datasets by analyzing and debiasing active learning algorithms. Active learning is broadly used for constructing large scale datasets. By studying biases in active learning algorithms, we attempt to reduce dataset biases which are root causes of biases in downstream applications.
Committee:
- Yanjun Qi, Committee Chair (Department of Computer Science)
- Vicente Ordóñez Román, Advisor (Department of Computer Science)
- Yangfeng Ji (Department of Computer Science)
- Jundong Li (Departments of Computer Science & Electrical and Computer Engineering, SEAS, DSI)
- Paul Humphreys (Department of Philosophy, GSAS, UVA)
- Olga Russakovsky (Department of Computer Science, Princeton University)