ZeroNAS: Differentiable Generative Adversarial Networks Search for Zero-Shot Learning

In recent years, remarkable progress in zero-shot learning (ZSL) has been achieved by generative adversarial networks (GAN). To compensate for the lack of training samples in ZSL, a surge of GAN architectures have been developed by human experts through trial-and-error testing. Despite their efficacy, however, there is still no guarantee that these hand-crafted models can consistently achieve good performance across diversified datasets or scenarios. Accordingly, in this paper, we turn to neural architecture search (NAS) and make the first attempt to bring NAS techniques into the ZSL realm. Specifically, we propose a differentiable GAN architecture search method over a specifically designed search space for zero-shot learning, referred to as ZeroNAS. Considering the relevance and balance of the generator and discriminator, ZeroNAS jointly searches their architectures in a min-max player game via adversarial training. Extensive experiments conducted on four widely used benchmark datasets demonstrate that ZeroNAS is capable of discovering desirable architectures that perform favorably against state-of-the-art ZSL and generalized zero-shot learning (GZSL) approaches. Source code is at https://github.com/caixiay/ZeroNAS.

1 Introduction

Generative Adversarial Networks (GANs) have shown promising results in generating data that are indistinguishable from real data [1], [2], [3]. Recently, a trend has emerged of synthesizing Convolutional Neural Network (CNN) features using GAN architectures, which mitigates the lack of unseen samples in zero-shot learning (ZSL) [4], [5], [6]. Of these methods, f-CLSWGAN [4] is one of the first attempts to leverage GANs in order to push the ZSL performance forward. In an attempt to progress this field, some improved approaches (e.g., LisGAN [7] and AFC-GAN [8]) that may potentially offer better performance, have been proposed. However, despite the empirical success of these approaches, it should be noted that they all rely heavily on hand-crafted GAN architectures designed by human experts, meaning that laborious trial-and-error testing is required (Fig. 1a). The instability issue in GAN training increases the difficulty of architecture design significantly. Once obtained, these manually designed architectures are fixed across all diversified data samples and application scenarios, which can easily lead to sub-optimal results. It is therefore highly valuable to automatically determine the GAN architectures customized for each specific ZSL task, rather than simply adopting a hand-crafted architecture. Fig. 1.

Comparison of the architecture design and training method of (a) Existing GANs for ZSL, (b) AutoGAN, (c) ZeroNAS.