QC-ODKLA: Quantized and Communication- Censored Online Decentralized Kernel Learning via Linearized ADMM

Area of Science:

• Machine Learning
• Decentralized Systems
• Signal Processing

Background:

• Online kernel learning enables agents to learn from streaming data collaboratively.
• Decentralized networks face challenges like high dimensionality and communication overhead.
• Reproducing Kernel Hilbert Space (RKHS) is crucial for nonlinear function approximation.

Purpose of the Study:

• To develop an efficient online kernel learning framework for decentralized networks.
• To address the curse of dimensionality using random feature mapping.
• To improve communication efficiency in decentralized learning algorithms.

Main Methods:

• Utilized random feature (RF) mapping to transform kernel learning into a parametric problem.
• Proposed the online decentralized kernel learning via linearized ADMM (ODKLA) framework.
• Introduced quantization and censoring for communication efficiency, creating QC-ODKLA.

Main Results:

• Theoretically proved optimal sublinear regret for both ODKLA and QC-ODKLA algorithms.
• Demonstrated effective learning, communication, and computation efficiency through numerical experiments.
• Validated the performance of the proposed decentralized kernel learning methods.

Conclusions:

• ODKLA and QC-ODKLA provide efficient solutions for online kernel learning in decentralized settings.
• The methods effectively manage dimensionality and communication constraints.
• The proposed algorithms offer a robust approach for collaborative learning from streaming data.