Unleashing the Power of Continual Learning on Non-Centralized Devices: A Survey

TL;DR

This survey presents a comprehensive overview of Non-Centralized Continual Learning (NCCL), addressing catastrophic forgetting, distribution shifts, heterogeneity, privacy, and real-world deployment on distributed devices. It systematically categorizes NCCL into three learning paradigms (DL, FL, HL) and three continual-learning methods (rehearsal, regularization, isolation), and articulates three NCCL protocol types (Task-IL, Domain-IL, Class-IL). The authors introduce data-, model-, and device-level strategies, discuss heterogeneity and security/privacy concerns, and provide a large-scale benchmark evaluating Federated Continual Learning approaches across diverse downstream tasks. They also explore real-world applications, resource constraints, and future directions, highlighting the growing role of LLM foundations and generative/replay-based techniques in NCCL. Overall, the paper consolidates current NCCL progress, benchmarks, and challenges, offering practical guidance and a roadmap for future research and deployment.

Abstract

Non-Centralized Continual Learning (NCCL) has become an emerging paradigm for enabling distributed devices such as vehicles and servers to handle streaming data from a joint non-stationary environment. To achieve high reliability and scalability in deploying this paradigm in distributed systems, it is essential to conquer challenges stemming from both spatial and temporal dimensions, manifesting as distribution shifts, catastrophic forgetting, heterogeneity, and privacy issues. This survey focuses on a comprehensive examination of the development of the non-centralized continual learning algorithms and the real-world deployment across distributed devices. We begin with an introduction to the background and fundamentals of non-centralized learning and continual learning. Then, we review existing solutions from three levels to represent how existing techniques alleviate the catastrophic forgetting and distribution shift. Additionally, we delve into the various types of heterogeneity issues, security, and privacy attributes, as well as real-world applications across three prevalent scenarios. Furthermore, we establish a large-scale benchmark to revisit this problem and analyze the performance of the state-of-the-art NCCL approaches. Finally, we discuss the important challenges and future research directions in NCCL.

Figures (11)

• Figure 1: The overview of NCCL paradigms, including the distributed device layer and centralized server layer with the continual learning process and three different NCL learning paradigms. Each device will collect new data during the training process and collaboratively maintain the model on all streaming tasks without breaching data privacy.
• Figure 2: The outline of this survey, where we introduce the provisioning of continual learning in the non-centralized learning paradigm and highlight some essential issues and challenges.
• Figure 3: Three typical Non-Centralized Learning paradigms include (a) Decentralized learning without the central server; (b) Federated learning with a central server aggregating clients; (c) Hierarchical learning with the client-edge-cloud model.
• Figure 4: Three Non-Centralized CL protocols. (a) represents the Task-IL scenario, where the boundaries of different tasks (e.g., task-id) are well-defined; (b) is the Domain-IL scenario, where there exists a feature shift among different tasks, but generally no new classes are introduced; (c) is the Class-IL scenario, where the class types of different tasks do not overlap.
• Figure 5: Data-level methods, including experience replay, generative replay, and proxy dataset. Experience replay focuses on reusing data and features stored in a local buffer; generative replay often relies on generative models to synthesize pseudo-data for replay; and a proxy dataset serves as a medium to facilitate knowledge transfer between different tasks.