Bootstrapped Unseen Prototypes for Few-Shot Inspired Generative Zero-Shot Learning

Md Shakil Ahamed Shohag, Q. M. Jonathan Wu, Ashab Uddin, Ning Zhang

Keywords: Zero-shot learning, generalized zero-shot learning, knowledge transfer, feature synthesis

Abstract: Generative zero-shot learning (ZSL) synthesizes visual features for unseen classes from semantic descriptors, then trains a fully supervised classifier. Although effective, most methods depend on large volumes of synthetic data and heavyweight generators which dilutes the original ZSL premise. We propose BUP-FSIGenZ framework that approaches the generative ZSL by taking few-shot learning (FSL) as an inspiration instead of conventional supervised formulation. Consequently, the method focuses on generating only a handful of bootstrapped prototypes per unseen class. Instead of modeling full distributions with adversarial or variational generators, we expose the variability of seen classes using statistical resampling and estimate the same for unseen classes by knowledge transfer to unseen domain. Concretely, we bootstrap the seen data to obtain multiple class prototypes that capture stable yet diverse modes; and then estimate unseen bootstrapped prototypes through knowledge transfer from seen to unseen domain. This way we generate few prototypes for each unseen class and use them as unseen synthetic training data. For classification, we introduce a classifier trained jointly with binary cross-entropy and KL-divergence objectives on visual–semantic contrast. This unified design drastically reduces compute and sample count, and attains competitive ZSL performance on SUN, AWA2, and CUB with significantly fewer synthetic features than conventional generative baselines.

Submission Number: 118