ADAMT: Adaptive distributed multi-task learning for efficient image recognition in Mobile Ad-hoc Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Mobile Computing

Background:

• Distributed machine learning in mobile ad hoc networks faces challenges from limited resources, non-IID data, and dynamic topologies.
• Existing methods often require centralized coordination and stable networks, limiting practical application.
• Efficient image recognition is crucial for edge devices in mobile environments.

Purpose of the Study:

• To propose ADAMT, an adaptive distributed multi-task learning framework for efficient image recognition in resource-constrained mobile ad hoc networks.
• To address limitations of existing distributed learning approaches in dynamic and resource-limited settings.
• To enable personalized model training and collaborative parameter updates on edge devices.

Main Methods:

• ADAMT employs a feature expansion mechanism for enhanced local model expressiveness.
• A deep hashing technique facilitates efficient on-device retrieval and multi-task fusion.
• An adaptive communication strategy dynamically adjusts model updates based on network conditions and node reliability.

Main Results:

• ADAMT achieved a top-1 accuracy of 0.867 on the ImageNet dataset, outperforming state-of-the-art methods.
• The framework significantly reduced communication overhead and accelerated convergence by 2.69 times compared to distributed SGD.
• The adaptive communication strategy effectively balanced model performance and resource consumption.

Conclusions:

• ADAMT offers an efficient and robust solution for distributed learning in mobile ad hoc networks.
• The framework is well-suited for resource-constrained environments, enabling advanced machine learning on edge devices.
• This work advances the design of distributed learning algorithms for practical edge AI applications.