Zeroshot

Zeroshot, often written zero-shot, refers to the ability of a model to perform a task or recognize classes without any task-specific training examples. In machine learning and artificial intelligence, zeroshot learning aims to generalize to unseen categories or tasks by leveraging auxiliary information such as semantic attributes, word or sentence embeddings, or natural language descriptions that relate seen and unseen concepts. In natural language processing and the use of large language models, zero-shot performance is achieved by prompts that specify the task without providing task-specific examples.

In computer vision, zeroshot learning maps visual representations to a semantic space and uses compatibility scores to assign an instance to an unseen class. Techniques include attribute-based models that describe classes with predefined attributes, embedding-based methods that learn a joint space for images and semantics, and generative approaches that synthesize feature representations for unseen classes. In natural language processing and multimodal systems, zero-shot capabilities emerge from prompting or architecture that leverage learned general knowledge, enabling tasks without labeled data for the target domain.

Zero-shot is distinct from few-shot learning, where a small number of labeled examples for the target task are provided. Common applications include image classification with unseen categories, cross-domain transfer, zero-shot transfer in robotics, and zero-shot evaluation of multimodal systems. Challenges include the semantic gap between the provided descriptions and real data, biases toward concepts with stronger representations, and scalability to large or nuanced concept sets. Evaluation protocols vary and can be sensitive to the chosen semantic representations, prompts, or embedding spaces.