Zero-shot Learning

Zero-shot learning refers to the ability and methods by which an AI model can perform inference and classification on classes or tasks it has never encountered during training. This makes it possible to build flexible AI systems that can handle categories or instructions not prepared in advance. In conventional machine learning, models are trained on data belonging to specific classes and then used to classify new data within those same classes. Zero-shot learning, by contrast, can reason about unknown classes by leveraging linguistic, conceptual knowledge and structure — without any explicit training on those classes. Key domains where zero-shot learning is applied: • Natural language processing (NLP): Large language models like GPT and T5 can handle diverse tasks (translation, summarization, classification, etc.) without explicit task-specific instructions • Image recognition: Multimodal models like CLIP learn correspondences between images and text, enabling classification of unseen image categories • Text classification, sentiment analysis, and Q&A: Tasks can be handled without specialized data by providing a natural language description of the task How zero-shot learning works: • Leverages knowledge encoded in pre-trained language or vision models • Takes task descriptions or goals as natural language input (prompts) • Infers relationships between unknown and known classes through embedding spaces and semantic representations Advantages of zero-shot learning: • No additional training required for new tasks or categories • Enables multilingual and multi-purpose applications (e.g., a single model handles multiple languages and formats) • Reduces annotation costs and improves development efficiency Alongside "few-shot learning" and "one-shot learning" — where a small number of examples are provided — zero-shot learning is gaining attention as a key technology toward building general-purpose AI. Zero-shot learning is a foundational approach toward AI that can "understand and apply meaning even without prior exposure" — much like human intelligence.