Tf-GCZSL: Task-free generalized continual zero-shot learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual zero-shot learning (CZSL) enables models to learn new classes over time without forgetting previous knowledge.
• Existing CZSL methods often require explicit task-boundary information during training, which is impractical for real-world scenarios.
• Catastrophic forgetting remains a significant challenge in continual learning frameworks.

Purpose of the Study:

• To propose a novel task-free generalized continual zero-shot learning (Tf-GCZSL) method.
• To eliminate the need for task-boundary information during the training of CZSL models.
• To enhance the model's ability to learn unseen classes while mitigating catastrophic forgetting.

Main Methods:

• A variational autoencoder (VAE) is employed for core zero-shot learning (ZSL) tasks.
• Short-term and long-term memory mechanisms are integrated to handle the absence of task boundaries.
• Experience replay, dark knowledge distillation, and regularization are combined to combat catastrophic forgetting.

Main Results:

• The proposed Tf-GCZSL method demonstrates significant performance improvements across five benchmark datasets (CUB, aPY, AWA1, AWA2, SUN).
• Improvements in prediction accuracy range from 1% to 12% compared to state-of-the-art and baseline methods.
• The method achieves strong results in both task-agnostic prediction and the proposed task-free settings.

Conclusions:

• The developed Tf-GCZSL method effectively addresses the limitations of existing CZSL approaches by removing the need for task-boundary information.
• The integration of memory, experience replay, and knowledge distillation proves effective in preventing catastrophic forgetting.
• The proposed approach offers a more practical and robust solution for continual learning in zero-shot scenarios.