A Generalization Theory for Zero-Shot Prediction
This work provides a theoretical foundation for zero-shot prediction in machine learning and AI, addressing a core challenge in modern generalization paradigms.
The paper tackles the problem of understanding generalization in zero-shot prediction, where pre-trained foundation models are used for downstream tasks without labeled data, by presenting a theoretical framework that identifies target quantities and conditional independence relationships enabling generalization.
A modern paradigm for generalization in machine learning and AI consists of pre-training a task-agnostic foundation model, generally obtained using self-supervised and multimodal contrastive learning. The resulting representations can be used for prediction on a downstream task for which no labeled data is available. We present a theoretical framework to better understand this approach, called zero-shot prediction. We identify the target quantities that zero-shot prediction aims to learn, or learns in passing, and the key conditional independence relationships that enable its generalization ability.