Zero-Shot Classification
Classifying images into categories the model has never seen during training - by reasoning about semantic relationships.
Think of it like
You know what a "dog" is and what "four-legged" means. Someone shows you a "dachshund" - youve never seen that breed, but you know its a four-legged dog.
Example
CLIP is trained on image-caption pairs. You ask it to classify a photo as one of ["cat", "dog", "elephant"] without ever seeing those three labels in training. It maps the image to embeddings and compares to the class name embeddings.
How it actually works
Bridges the train-test gap by using semantic attributes or language. CLIP and similar models encode images and text into a shared embedding space; classification is a nearest-neighbor lookup. Works for unseen classes by leveraging language understanding. Limited by coverage of the text space and spurious correlations in the joint embedding.
For product teams
Enables systems to classify novel categories without retraining - huge for scalability and rapid iteration.
For engineers
Models: CLIP, Flamingo. Metrics: accuracy on held-out classes. Datasets test on unseen classes: ImageNet-21K variants.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome