PAMK: Prototype Augmented Multi-Teacher Knowledge Transfer Network for Continual Zero-Shot Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual zero-shot learning (CZSL) aims to learn new tasks without forgetting previous ones.
• Existing CZSL methods often suffer from negative transfer, hindering new task generalization.
• This is due to over-reliance on old knowledge and reduced model plasticity.

Purpose of the Study:

• To propose PAMK, a prototype augmented multi-teacher knowledge transfer network for CZSL.
• To balance recognition stability for old tasks and generalization plasticity for new tasks.
• To overcome catastrophic forgetting and the negative transfer problem in CZSL.

Main Methods:

• PAMK utilizes a prototype augmented contrastive generation (PACG) module and a multi-teacher knowledge transfer (MKT) module.
• PACG employs a continual prototype augmentation strategy and a semantic-visual contrastive loss.
• MKT facilitates knowledge accumulation from old tasks to new tasks via multi-teacher transfer.

Main Results:

• PAMK demonstrates superior performance over state-of-the-art methods in various CZSL settings.
• Significant gains in mean harmonic accuracy were observed in the task-free CZSL setting.
• Improvements of 3.28%, 3.09%, and 3.71% were achieved on CUB, AWA1, and AWA2 datasets, respectively.

Conclusions:

• PAMK effectively mitigates catastrophic forgetting and the negative transfer problem in CZSL.
• The proposed method enhances both stability for old tasks and plasticity for new tasks.
• PAMK offers a promising solution for robust and adaptable continual learning systems.