AI and ML models are trained on large datasets that contain labelled examples classified in a range of categories. But what if AI could make predictions and recognize objects belonging to  unseen categories or classes?  Applying AI in a more generic way, above and beyond the data it has been trained on , is what Zero Shot Learning (ZSL) is all about.

ZSL is a new paradigm. It enables the model to classify or recognize objects it has never encountered during training. It doesn’t rely on data label pairs. It makes use of auxiliary information available such as text descriptions and semantic attributes. The model learns the relationship between an object and the terms (metadata) that are used to describe it.

Quality metadata with comprehensive descriptions is what ZSL hinges on entirely. Poor descriptions and semantic attributes will affect the efficacy of ZSL. Another challenge is evaluating ZSL’s ability to generalize, which in the real world will not be easy to do with standard accuracy metrics used to measure the results of other models.

ZSL represents a shift in how AI can reason and is a distinct approach towards more generalized AI that doesn’t entirely rely on massive data labelling. ZSL in combination with with few-show learning can deliver far better accuracy and flexibility in the future.