Adaptive information-constrained mapping for feature compression in edge AI and federated systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Compression

Background:

• Distributed intelligent systems face challenges with limited resources, impacting Edge AI and Federated Learning.
• Reducing communication overhead is crucial due to unstable Quality of Service, limited bandwidth, and data heterogeneity.

Purpose of the Study:

• To develop an efficient feature compression model for resource-constrained distributed intelligent systems.
• To address the need for reduced communication overhead in Edge AI and Federated Learning environments.

Main Methods:

• Developed a novel entropy-regularised compression model combining variational latent mapping, non-negativity-constrained projection, and stochastic-Boolean transformation.
• Proposed a generalised compression quality functional incorporating Kullback-Leibler divergence and semantic relevance preservation.
• Designed efficient projection-gradient optimisation algorithms for constrained computational environments.

Main Results:

• Achieved a 6-fold reduction in entropy load on HAR and PAMAP2 datasets.
• Maintained classification accuracy above 94% with high semantic fidelity.
• Demonstrated robustness to noise and loss on low-power devices (Jetson Nano, Raspberry Pi 4).

Conclusions:

• The proposed model offers superior compression efficiency, adaptability, and stability compared to SOTA solutions.
• The approach is effective for Edge AI and Federated Learning, especially under unstable transmission conditions.
• Validated practical effectiveness on real-world datasets and low-power hardware.