Class-Incremental Cloud-Device Collaborative Adaptation With Contrastive Learning in Dynamic Changing Environments.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Lightweight models are crucial for edge devices, but struggle with generalization in dynamic environments.
• Cloud-Device Collaborative Learning (CDCL) transfers knowledge from cloud to edge, yet current methods neglect continually incoming new classes.
• Adapting Class-Incremental Learning (CIL) to CDCL faces challenges like poor lightweight model generalization and data-incremental learning overfitting.

Purpose of the Study:

• To address the limitations of existing CDCL methods in handling continual new classes.
• To propose a novel framework for class-incremental cloud-device collaborative learning (CI-CDCL).
• To enhance the generalization and reduce overfitting of lightweight models in dynamic, incremental learning scenarios.

Main Methods:

• A novel contrastive prototypical network based CI-CDCL framework is proposed.
• The framework aims to improve collaborative learning effectiveness for both class-incremental and data-incremental learning.
• Extensive experiments were conducted on public datasets to validate the model's performance.

Main Results:

• The proposed CI-CDCL framework demonstrates effectiveness in improving cloud-device collaboration.
• The model addresses key challenges of poor generalization and overfitting in incremental learning settings.
• Experimental results validate the proposed approach on benchmark datasets.

Conclusions:

• The developed CI-CDCL framework offers a promising solution for continual learning on edge devices.
• This work advances CDCL by incorporating class-incremental learning capabilities.
• The contrastive prototypical network approach enhances the adaptability and robustness of lightweight models.