Decentralized Learning

Communication-efficient distributed machine learning for edge computing

Decentralized Learning for Edge Computing Applications

The increasing deployment of edge devices and the need for privacy-preserving machine learning have made decentralized learning an important research area. Our work focuses on developing communication-efficient algorithms for training deep learning models across distributed nodes without requiring a central coordinator.

DIMAT: Decentralized Iterative Merging-and-Training (2024)

Our latest work, DIMAT (Decentralized Iterative Merging-and-Training), provides a novel framework for efficient decentralized deep learning:

  • Iterative model merging with local training phases
  • Theoretical convergence guarantees
  • Adaptability to various network topologies
  • Significant reduction in communication overhead
DIMAT Architecture
Architectural overview of our Decentralized Iterative Merging-and-Training (DIMAT) approach presented at CVPR 2024.

Cross-Gradient Aggregation for Non-IID Data (2021)

A significant challenge in decentralized learning is dealing with non-Independent and Identically Distributed (non-IID) data. We’ve addressed this with Cross-Gradient Aggregation:

  • Handles data heterogeneity across distributed nodes
  • Maintains model accuracy despite data distribution differences
  • Supports personalization while enabling collaborative learning
Cross-Gradient Aggregation
Convergence Comparison
Left: Visualization of our Cross-Gradient Aggregation approach for non-IID data. Right: Convergence comparison with baseline decentralized learning methods.

Momentum-Based Decentralized Policy Gradient Tracking (2022)

We have developed MDPGT (Momentum-based Decentralized Policy Gradient Tracking), a novel algorithm for reinforcement learning in decentralized settings:

  • Communication efficiency: Reduces network bandwidth requirements
  • Faster convergence: Leverages momentum to accelerate training
  • Robustness: Maintains performance even with unreliable network connections
MDPGT Architecture
Architectural overview of our Momentum-based Decentralized Policy Gradient Tracking (MDPGT) approach.

Publications

Key publications in this research area include:

  1. Saadati, N., Pham, M., Saleem, N., Waite, J.R., Balu, A., Jiang, Z., Hegde, C., & Sarkar, S. (2024). DIMAT: Decentralized Iterative Merging-and-Training for Deep Learning Models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 27517-27527.

  2. Jiang, Z., Lee, X.Y., Tan, S.Y., Tan, K.L., Balu, A., Lee, Y.M., Hegde, C., & Sarkar, S. (2022). MDPGT: Momentum-based Decentralized Policy Gradient Tracking. Proceeding of AAAI Conference on Artificial Intelligence.

  3. Esfandiari, Y., Tan, S.Y., Jiang, Z., Balu, A., Herron, E., Hegde, C., & Sarkar, S. (2021). Cross-gradient Aggregation for Decentralized Learning from Non-IID Data. Proceedings of the 38th International Conference on Machine Learning, 3036-3046.

Applications and Future Work

Our decentralized learning research has applications in:

  • Agricultural IoT networks
  • Federated medical image analysis
  • Privacy-preserving collaborative AI
  • Edge computing for autonomous systems

We are currently extending this work to incorporate differential privacy guarantees, adaptive communication protocols, and heterogeneous device capabilities.