Scalable Zero-Shot Learning via Binary Visual-Semantic Embeddings.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to classify visual instances from unseen classes without direct training examples.
• Current ZSL methods often map visual features to semantic spaces, which can be unreliable due to noisy embeddings or visual bias.
• A robust method is needed to address the visual-semantic bias inherent in traditional ZSL approaches.

Purpose of the Study:

• To propose a novel Binary embedding based Zero-Shot Learning (BZSL) method.
• To alleviate the visual-semantic bias problem in zero-shot learning.
• To achieve efficient training and scalable inference for unseen classes.

Main Methods:

• Developed a Binary embedding based Zero-Shot Learning (BZSL) method.
• Jointly learned two binary coding functions to encode visual instances and class embeddings into a discriminative Hamming space.
• Formulated an equivalent correlation maximization problem with an analytical solution for efficient training.

Main Results:

• The BZSL method effectively alleviates the visual-semantic bias problem.
• Achieved superior performance on zero-shot learning tasks across four benchmark datasets, including ImageNet.
• Demonstrated that increasing binary embedding dimensions improves recognition accuracy.

Conclusions:

• BZSL offers a novel and effective approach for zero-shot learning by utilizing a discriminative Hamming space.
• The proposed method provides an efficient training process and scalable inference for novel classes.
• Binary embedding dimension is a critical factor for enhancing zero-shot recognition accuracy.