CAST: Closed-form Analytic Semantic Transfer for Zero-Shot Classifier Extension

2026-08-13

AI Summary

This study addresses the absence of target samples in zero-shot learning by proposing a training-free, image-free analytical semantic transfer framework. Through a closed-form weight injection mechanism, pretrained classifiers are extended to unseen categories without iterative optimization. Furthermore, this work establishes a finite-sample error decomposition theory and introduces a computable semantic extrapolation residual metric to guide data construction. Empirical evaluations on standard benchmarks demonstrate that the proposed method matches or surpasses existing image-free approaches while approaching few-shot performance levels. Collectively, this framework provides an efficient analytical solution with rigorous theoretical support for zero-shot learning, effectively bridging the gap between training-free efficiency and high-accuracy generalization in the absence of visual data.

Abstract

Large pre-trained models have become foundational components of modern machine learning systems. Yet adapting these models to novel categories typically requires examples from the target distribution. In many domains, however, such data are unavailable. Zero-shot learning (ZSL) permits recognition under these limitations through relying on auxiliary semantic information such as textual descriptions. We introduce CAST (Closed-form Analytic Semantic Transfer), a training-free, image-free framework for extending a pre-trained classifier to previously unseen classes through weight injection. We provide a theoretical foundation for CAST and derive a finite-sample error decomposition that identifies the \emph{semantic extrapolation residual} $ρ_u$. The residual is a computable, model-agnostic measure and provides a principled criterion for dataset curation and benchmark design. Experiments on standard zero-shot learning benchmarks demonstrate that CAST matches or exceeds existing image-free approaches and approaches the performance of few-shot adaptation methods, while requiring neither iterative optimization nor examples from the target distribution.

Problem

• Zero-Shot Learning
• Classifier Extension
• Pre-trained Models
• Semantic Transfer

Innovation

• Zero-Shot Learning
• Training-Free
• Weight Injection
• Semantic Extrapolation Residual
• Closed-Form Analytic