Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FedEMA: Federated Exponential Moving Averaging with Negative Entropy Regularizer in Autonomous Driving

Wei-Bin Kou, Guangxu Zhu, Bingyang Cheng, Shuai Wang, Ming Tang, Yik-Chung Wu

TL;DR

The paper addresses temporal forgetting and limited generalization in federated learning for autonomous driving S3U tasks under evolving environments. It introduces FedEMA, combining a server-side EMA fusion with a vehicle-side negative entropy regularizer to balance stability and adaptability, supported by convergence analysis showing a $O(1/\sqrt{R})$ rate and practical gains. The method demonstrates superior performance over state-of-the-art FL baselines on Cityscapes and CamVid across DeepLabv3+ and TopFormer backbones, with notable improvements in mean IoU and related metrics. This approach offers a practical path to robust, privacy-preserving domain-generalization for AD systems in dynamic real-world settings.

Abstract

Street Scene Semantic Understanding (denoted as S3U) is a crucial but complex task for autonomous driving (AD) vehicles. Their inference models typically face poor generalization due to domain-shift. Federated Learning (FL) has emerged as a promising paradigm for enhancing the generalization of AD models through privacy-preserving distributed learning. However, these FL AD models face significant temporal catastrophic forgetting when deployed in dynamically evolving environments, where continuous adaptation causes abrupt erosion of historical knowledge. This paper proposes Federated Exponential Moving Average (FedEMA), a novel framework that addresses this challenge through two integral innovations: (I) Server-side model's historical fitting capability preservation via fusing current FL round's aggregation model and a proposed previous FL round's exponential moving average (EMA) model; (II) Vehicle-side negative entropy regularization to prevent FL models' possible overfitting to EMA-introduced temporal patterns. Above two strategies empower FedEMA a dual-objective optimization that balances model generalization and adaptability. In addition, we conduct theoretical convergence analysis for the proposed FedEMA. Extensive experiments both on Cityscapes dataset and Camvid dataset demonstrate FedEMA's superiority over existing approaches, showing 7.12% higher mean Intersection-over-Union (mIoU).

FedEMA: Federated Exponential Moving Averaging with Negative Entropy Regularizer in Autonomous Driving

TL;DR

The paper addresses temporal forgetting and limited generalization in federated learning for autonomous driving S3U tasks under evolving environments. It introduces FedEMA, combining a server-side EMA fusion with a vehicle-side negative entropy regularizer to balance stability and adaptability, supported by convergence analysis showing a rate and practical gains. The method demonstrates superior performance over state-of-the-art FL baselines on Cityscapes and CamVid across DeepLabv3+ and TopFormer backbones, with notable improvements in mean IoU and related metrics. This approach offers a practical path to robust, privacy-preserving domain-generalization for AD systems in dynamic real-world settings.

Abstract

Street Scene Semantic Understanding (denoted as S3U) is a crucial but complex task for autonomous driving (AD) vehicles. Their inference models typically face poor generalization due to domain-shift. Federated Learning (FL) has emerged as a promising paradigm for enhancing the generalization of AD models through privacy-preserving distributed learning. However, these FL AD models face significant temporal catastrophic forgetting when deployed in dynamically evolving environments, where continuous adaptation causes abrupt erosion of historical knowledge. This paper proposes Federated Exponential Moving Average (FedEMA), a novel framework that addresses this challenge through two integral innovations: (I) Server-side model's historical fitting capability preservation via fusing current FL round's aggregation model and a proposed previous FL round's exponential moving average (EMA) model; (II) Vehicle-side negative entropy regularization to prevent FL models' possible overfitting to EMA-introduced temporal patterns. Above two strategies empower FedEMA a dual-objective optimization that balances model generalization and adaptability. In addition, we conduct theoretical convergence analysis for the proposed FedEMA. Extensive experiments both on Cityscapes dataset and Camvid dataset demonstrate FedEMA's superiority over existing approaches, showing 7.12% higher mean Intersection-over-Union (mIoU).
Paper Structure (20 sections, 1 theorem, 11 equations, 5 figures, 6 tables, 1 algorithm)

This paper contains 20 sections, 1 theorem, 11 equations, 5 figures, 6 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Federated Autonomous Driving (FedAD) System
  4. Street Scene Semantic Understanding (S3U)
  5. Methodology
  6. FedAD Formulation
  7. Vehicle Local Training
  8. Server Aggregation
  9. The Proposed FedEMA
  10. FedEMA Convergence Analysis
  11. Experiments
  12. Datasets, Evaluation Metrics, and Implementation
  13. Datasets
  14. Evaluation Metrics
  15. Implementation Details
  16. ...and 5 more sections

Key Result

Theorem 1

Under non-convex loss setting, after $R$ rounds' training with learning rate $\eta = \frac{1}{\sqrt{R}}$, we have where $\mathcal{L}^*$ is the theoretical optimal loss and can be treated as a constant, and $\omega^0$ is the initial model of the training (i.e., $\omega^0 = \omega^0_{\text{EMA}}$).

Figures (5)

  • Figure 1: Illustration of Federated AD (FedAD) system kou2024pfedlvm. The inherent training scheme of FedAD, which continuously overwrites old local model with new global model, inevitably leads to the catastrophic forgetting problem, particularly as the dataset is dynamic over time.
  • Figure 2: Overview of the proposed FedEMA method.
  • Figure 3: Metric comparison of DeepLabv3+ on Cityscapes dataset. Results show FedEMA outperform other FL methods across all metrics.
  • Figure 4: Performance and convergence comparison with different regularization coefficient $\lambda$ while keeping window size $N=5$.
  • Figure 5: Performance and convergence comparison with different window size $N$ while keeping regularization coefficient $\lambda=0.002$.

Theorems & Definitions (1)

  • Theorem 1