A survey and experimental study for embedding-aware generative models: Features, models, and any-shot scenarios

Abstract

In the era of industrial artificial intelligence, grappling with data insufficiency remains a formidable challenge that stands at the forefront of our progress. The embedding-aware generative model emerges as a promising solution, tackling this issue head-on in the realm of zero-shot learning by ingeniously constructing a generator that bridges the gap between semantic and feature spaces. Thanks to the predefined benchmark and protocols, the number of proposed embedding-aware generative models for zero-shot learning is increasing rapidly. We argue that it is time to take a step back and reconsider the embedding-aware generative paradigm. The main work of this paper is two-fold. First, embedding features in benchmark datasets are somehow overlooked, which potentially limits the performance of generative models, while most researchers focus on how to improve them. Therefore, we conduct a systematic evaluation of 10 representative embedding-aware generative models and prove that even simple representation modifications on the embedding features can improve the performance of generative models for zero-shot learning remarkably. So it is time to pay more attention to the current embedding features in benchmark datasets. Second, based on five benchmark datasets, each with six any-shot learning scenarios, we systematically compare the performance of ten typical embedding-aware generative models for the first time, and we give a strong baseline for zero-shot learning and few-shot learning. Meanwhile, a comprehensive generative model repository, namely, generative any-shot learning repository, is provided, which contains the models, features, parameters, and scenarios of embedding-aware generative models for zero-shot learning and few-shot learning. Any results in this paper can be readily reproduced with only one command line based on generative any-shot learning.