Zero-Shot Learning

Published:

Zero-shot learning is when a model can handle something new without being trained on labeled examples for that exact case. Instead of learning the new category from examples, it uses a description of what the category means and connects it to knowledge it already has. That is why a system can sometimes recognize an unseen label from its name or follow a new instruction without needing task-specific training data.

This works best when the description gives a clear link to what the model has learned before. If the label is vague or the description is thin, the model may map it to the wrong concept even while sounding confident. Clear wording and well-defined labels make a big difference in how reliable zero-shot behavior is.

Follow us on Facebook and LinkedIn to keep abreast of our latest news and articles