Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Model Pruning Enables Efficient Federated Learning on Edge Devices (1909.12326v5)

Published 26 Sep 2019 in cs.LG, cs.DC, and stat.ML

Abstract: Federated learning (FL) allows model training from local data collected by edge/mobile devices while preserving data privacy, which has wide applicability to image and vision applications. A challenge is that client devices in FL usually have much more limited computation and communication resources compared to servers in a datacenter. To overcome this challenge, we propose PruneFL -- a novel FL approach with adaptive and distributed parameter pruning, which adapts the model size during FL to reduce both communication and computation overhead and minimize the overall training time, while maintaining a similar accuracy as the original model. PruneFL includes initial pruning at a selected client and further pruning as part of the FL process. The model size is adapted during this process, which includes maximizing the approximate empirical risk reduction divided by the time of one FL round. Our experiments with various datasets on edge devices (e.g., Raspberry Pi) show that: (i) we significantly reduce the training time compared to conventional FL and various other pruning-based methods; (ii) the pruned model with automatically determined size converges to an accuracy that is very similar to the original model, and it is also a lottery ticket of the original model.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Yuang Jiang (12 papers)
  2. Shiqiang Wang (79 papers)
  3. Victor Valls (43 papers)
  4. Bong Jun Ko (5 papers)
  5. Wei-Han Lee (11 papers)
  6. Kin K. Leung (65 papers)
  7. Leandros Tassiulas (89 papers)
Citations (369)

Summary

An Overview of "Model Pruning Enables Efficient Federated Learning on Edge Devices"

The paper "Model Pruning Enables Efficient Federated Learning on Edge Devices" addresses a significant challenge in the field of federated learning (FL): the computational and communication resource constraints faced by edge devices. Typically, FL involves training models on decentralized data across various client devices, which remain resource-constrained compared to servers in centralized data environments. This research introduces PruneFL, a federated learning approach designed to reduce the computational overhead on these devices through adaptive model pruning.

Key Contributions

The paper offers several key contributions to model pruning in federated learning:

  1. Two-Stage Distributed Pruning: The authors propose an initial pruning stage at a selected client followed by further pruning among all clients during the FL process. This two-tiered approach is particularly effective in situations where data and computational capabilities are non-uniformly distributed among clients.
  2. Adaptive Pruning: PruneFL continuously evaluates and adapts the model's size. This dynamic adaptation is crucial for maintaining a balance between model accuracy and the efficiency of training and communication. The method involves quantifying the empirical risk reduction against the time required per FL round, thus iteratively refining the optimal size of the model.
  3. Implementation on Edge Devices: A practical implementation of PruneFL is provided on Raspberry Pi devices, demonstrating that the system can run with limited resources by adapting neural network training processes to use sparse matrices efficiently. This implementation addresses the lack of support for effective sparse matrix operations in existing machine learning frameworks.

Experimental Results

The experiments in this paper validate the advantages of PruneFL compared to traditional federated learning and other pruning techniques. Here are some of the notable results:

  • Training Efficiency: PruneFL significantly reduces training time compared to conventional FL and iterative pruning, providing acceleration without losing model accuracy. For instance, the paper reports that the training time on Raspberry Pi devices sees a moderate reduction through the method's sparse matrix computations.
  • Model Accuracy: Despite the reductions in training time and model size, PruneFL achieves a similar convergence accuracy to the unpruned, full-sized models. It successfully achieves the balance of model size and accuracy, adapting automatically during the training process.
  • Utility as Lottery Tickets: The pruned models found by PruneFL potentially align with the "lottery ticket hypothesis," suggesting that these sub-networks are not only smaller but retain sufficient trainable capacity.

Implications and Future Directions

The implications of PruneFL are tangible in practical scenarios where federated learning is applied on edge devices, such as mobile phones or IoT devices, where computation efficiency is as critical as maintaining model performance. This method can significantly widen the applicability of FL in resource-constrained environments and showcase how dynamic model adaptation can mitigate overheads related to computation and communication.

Future research could expand on this work by exploring more sophisticated adaptive methods for pruning or investigating the integration of other compression techniques like quantization. Additionally, the findings in this paper are likely to stimulate further exploration into enabling efficient sparse matrix computation in popular machine learning frameworks over low-power devices, which remains an underexplored avenue largely limited by current software and hardware constraints.

Overall, this paper makes notable contributions to the field of efficient federated learning, offering a method to enhance the practicality and scalability of deploying federated learning systems across a diverse range of real-world edge devices.