Apprentissage zero-shot
Définition
Zero-shot learning is solving a task sans aucun exemple étiqueté pour cette tâche. les LLMs font cela via le prompting; vision models can do it with text-conditioned classifiers (par 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 (par 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.
Comment ça fonctionne
The task is described in natural language (par ex. prompt: “Classify the sentiment as positive or negative”) or via a shared representation (par ex. attribute vectors, text embeddings). Input (par ex. a sentence or image) is fed to the model along with the task description. The model produces an output (par 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.
Cas d'utilisation
Zero-shot learning fits when you want to run a task with no target-task training—only a task description (par ex. prompt or class names).
- LLM tasks via prompts (par ex. classification, summarization) without fine-tuning
- CLIP-style image classification from text descriptions
- New categories or languages without labeled examples