Zero-Shot Classification
The ability to classify data into categories the model has never been explicitly trained on, by leveraging general knowledge learned during pre-training.
How It Works with LLMs
Simply describe the categories in a prompt: 'Classify this text as positive, negative, or neutral: [text]'. The model uses its understanding of language to classify without any task-specific training examples.
How It Works with CLIP
CLIP embeds both images and text descriptions. To classify an image, encode candidate labels as text ('a photo of a dog', 'a photo of a cat') and find the most similar text embedding. No training on those specific classes needed.
Practical Impact
Zero-shot classification eliminates the need for labeled training data for new categories. This is transformative for applications where creating labeled datasets is expensive or where categories change frequently.