New CAST framework enables zero-shot classifier extension without target data

Researchers have introduced CAST, a novel framework designed to extend pre-trained classifiers to new categories without requiring any examples from the target distribution. This training-free and image-free approach utilizes auxiliary semantic information, such as textual descriptions, to enable recognition of unseen classes through weight injection. CAST is supported by a theoretical foundation that includes a computable, model-agnostic measure called the semantic extrapolation residual, which can guide dataset curation and benchmark design. Experiments show that CAST performs comparably to or better than existing image-free methods and approaches the effectiveness of few-shot adaptation techniques. AI

IMPACT Enables AI models to recognize new categories without needing specific training data, potentially broadening their applicability.

RANK_REASON Academic paper detailing a new method for zero-shot learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CAST
• CatalyzeX
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →