Aprendizado zero-shot
Definição
Zero-shot learning é resolver uma tarefa sem nenhum exemplo rotulado para essa tarefa. Os LLMs fazem isso via prompting; modelo de vision models can do it with text-conditioned classifiers (por ex. CLIP).
No fine-tuning or few-shot examples are used; the task is specified only by description or by mapping to a shared space (por ex. text). LLMs excel at zero-shot for many NLP tasks; CLIP and similar models enable zero-shot image classification from text. Quality depends on how well pretraining covered the task or similar ones.
Como funciona
The task is described in natural language (por ex. prompt: “Classify the sentiment as positive or negative”) or via a shared representation (por ex. attribute vectors, text embeddings). Input (por ex. a sentence or image) is fed to the model along with the task description. The model produces an output (por ex. label, summary) using only what it learned at pretraining—no gradient updates on the target task. For CLIP: image and text are embedded in a shared space; zero-shot classification is done by comparing the image embedding to class name embeddings. For LLMs: the prompt states the task and format; the model completes accordingly.
Casos de uso
Zero-shot learning fits when you want to run a task with no target-task training—only a task description (por ex. prompt or class names).
- LLM tasks via prompts (por ex. classification, summarization) without fine-tuning
- CLIP-style image classification from text descriptions
- New categories or languages without labeled examples