International Conference on Computer and Knowledge Engineering

Home / 12th International Conference on Computer and Knowledge Engineering

Averting Mode Collapse for Generative Zero-Shot Learning

Authors :

Shayan Ramazi¹ Setare Shabani²

1- Cyberspace research institute, Shahid Beheshti University, G.C., Tehran, Iran 2- Department of Information Technology, K. N. Toosi University of Technology, Tehran, Iran

Keywords :

Generative Zero-shot learning،Zero-shot learning(ZSL)،image classification،Generalized Zero-shot learning(GZSL)،Generative Adversarial Networks(GAN)

Abstract :

Zero-shot learning is a promising method for supervised learning methods when there is no labeled data from some classes. In fact, in zero-shot learning, the model classifies the data with no labeled sample present in the training phase. Recently, researchers have shown that the use of conditional adversarial generative networks can be a solution to generate synthetic data from classes that do not have labeled data. In this paper, we try to improve the quality and variety of data generated from classes with no labeled sample through an adversarial generative network by introducing a new architecture. In addition to the changes made in the training phase of the adversarial generator network, a regulator has also been utilized to ensure the diversity of generated data. This regulator also prevents mode collapse. The proposed architecture is assessed on the Animals With Attributes (AWA) dataset. Classification accuracy in ZSL and GZSL settings improved by 1.7% and 2.3% in comparison to the state-of-the-art results.