HFM: A Hybrid Feature Model Based on Conditional Auto Encoders for Zero-Shot Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep Learning (DL) models typically require extensive training data.
• Zero-Shot Learning (ZSL) addresses this by enabling classification of unseen classes not present in the training set.
• Existing ZSL methods face challenges in effectively utilizing limited or no training data for novel categories.

Purpose of the Study:

• To propose a novel Hybrid Feature Model (HFM) for Zero-Shot Learning (ZSL) and Generalized Zero-Shot Learning (GZSL).
• To overcome the data-hungry nature of Deep Learning models by generating synthetic training data for unseen classes.
• To improve classification performance in scenarios with limited or no direct training examples for target classes.

Main Methods:

• A Hybrid Feature Model (HFM) employing two conditional autoencoders is proposed.
• The first autoencoder is trained on visual and semantic spaces; the second on visual space alone.
• Decoders from both autoencoders generate pseudo-training data for unseen classes, which then trains a support vector machine.

Main Results:

• The HFM demonstrates promising performance on four benchmark datasets.
• The proposed method achieves competitive results compared to current state-of-the-art approaches in both ZSL and GZSL settings.
• Effective generation of pseudo-training data enables successful classification of previously unseen classes.

Conclusions:

• The Hybrid Feature Model (HFM) offers an effective solution for Zero-Shot Learning (ZSL) by leveraging conditional autoencoders.
• The approach successfully generates pseudo-training data, mitigating the need for large labeled datasets.
• The method shows significant potential for real-world applications requiring classification of novel categories.