A Probabilistic Zero-Shot Learning Method via Latent Nonnegative Prototype Synthesis of Unseen Classes

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to classify unseen classes without direct training data.
• Conventional ZSL methods struggle with performance degradation due to nonuniform variances between categories.
• Nearest neighbor search (NNS)-based methods are common but sensitive to variance issues.

Purpose of the Study:

• To propose a probabilistic framework for ZSL that incorporates covariance to address performance degradation.
• To define a new latent space that promotes intra-class compactness and inter-class separability.
• To synthesize prototypes for unseen classes using nonnegative matrix factorization (NMF).

Main Methods:

• A probabilistic framework is developed, considering covariance for improved ZSL.
• A novel latent space is defined using triplet learning for feature gathering and scattering.
• Nonnegative matrix factorization (NMF) synthesizes unseen class prototypes based on attribute relationships.
• Visual features are projected into the latent space for classification in both classic and generalized ZSL.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art approaches on four benchmark datasets.
• The framework effectively handles nonuniform variances between categories, a common challenge in ZSL.
• Experimental results validate the efficacy of the latent space and prototype synthesis approach.

Conclusions:

• The proposed probabilistic framework offers a robust solution for zero-shot learning by accounting for category covariance.
• The novel latent space and NMF-based prototype synthesis significantly improve classification accuracy for unseen classes.
• This approach advances the field of ZSL, particularly for generalized ZSL scenarios.