Zero-shot learning

Zero-shot learning is the ability to perform a task without any task-specific training examples. In the context of language models, zero-shot approaches leverage knowledge acquired during pretraining to solve new tasks described via natural language instructions. Zero-shot performance avoids the costs and data requirements of task-specific supervision while reducing the risk of spurious feature-label correlations.

Key papers

• Toxicity Detection with Generative Prompt-based Inference — Demonstrates zero-shot toxicity detection via prompt-based inference without any toxicity-labeled supervision

Related topics

• Prompt-based methods — how zero-shot tasks are specified and performed
• Few-Shot Learning — extending to small numbers of examples
• Language Models — the models enabling zero-shot generalization
• Transfer learning for fake news detection — broader notion of using pretraining for new tasks