Incremental Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional zero-shot learning (ZSL) struggles with incrementally collected data, limiting real-world applications.
• Existing ZSL methods are trained on fixed class sets and cannot adapt to evolving data streams.

Purpose of the Study:

• To introduce a novel Incremental Zero-Shot Learning (IZSL) setting for recognizing objects from new classes while retaining knowledge of previously learned classes.
• To develop a method that addresses catastrophic forgetting in ZSL by accumulating historical knowledge.

Main Methods:

• A generative replay strategy is used to create virtual samples of previously encountered classes.
• Joint training on new real data and virtual old data facilitates knowledge transfer.
• Knowledge distillation is employed to regularize the model and transfer knowledge from previous to current learning stages.

Main Results:

• The proposed method effectively handles the IZSL problem, outperforming existing ZSL approaches.
• Experiments on three benchmarks demonstrate the efficacy of the generative replay and knowledge distillation strategies.
• The method shows significant improvements in recognizing objects from incrementally learned classes.

Conclusions:

• The novel IZSL setting and proposed method offer a viable solution for real-world ZSL applications with continuous data streams.
• The approach successfully mitigates catastrophic forgetting, enabling robust incremental learning in ZSL.
• This work advances ZSL capabilities by enabling models to learn and adapt over time.