Continual Learning for Smart City: A Survey

TL;DR

This survey addresses the need for continual learning in smart cities by surveying foundational CL concepts, method classes, and evaluation metrics, then mapping them to urban applications across transportation, environment, health, safety, networks, and robotics. It categorizes three core CL families (regularization, replay, architecture) and discusses advanced frameworks that couple CL with graph, temporal, spatial-temporal, multi-modal, and federated learning. Key contributions include a comprehensive taxonomy of CL methods in urban contexts, a catalog of relevant open datasets, and a discussion of open challenges such as privacy, security, explainability, and open-world adaptation. The work aims to guide researchers and practitioners toward robust, scalable, and privacy-conscious continual learning solutions for dynamic urban environments.

Abstract

With the digitization of modern cities, large data volumes and powerful computational resources facilitate the rapid update of intelligent models deployed in smart cities. Continual learning (CL) is a novel machine learning paradigm that constantly updates models to adapt to changing environments, where the learning tasks, data, and distributions can vary over time. Our survey provides a comprehensive review of continual learning methods that are widely used in smart city development. The content consists of three parts: 1) Methodology-wise. We categorize a large number of basic CL methods and advanced CL frameworks in combination with other learning paradigms including graph learning, spatial-temporal learning, multi-modal learning, and federated learning. 2) Application-wise. We present numerous CL applications covering transportation, environment, public health, safety, networks, and associated datasets related to urban computing. 3) Challenges. We discuss current problems and challenges and envision several promising research directions. We believe this survey can help relevant researchers quickly familiarize themselves with the current state of continual learning research used in smart city development and direct them to future research trends.

Figures (8)

• Figure 1: An example of continual learning for graph learning. In (a), a Model is firstly trained based on the data from Task A and then updated by Task B as Model*. Often, the new data distribution in Task B can be OOD. If the catastrophic forgetting problem is not handled well enough, the latest Model* can have degenerated performance on the previous Task A, shown in (b).
• Figure 2: The trend of continual learning research and its application to smart city research. Statistics are from Google Scholar over the past five years.
• Figure 3: A taxonomy of common continual learning methods to overcome catastrophic forgetting. (a). Regularization-based methods use additional regularization terms to prevent the important model parameters from deviating too much during sequential training. (b). Replay-based methods maintain a memory buffer or a generative model to replay the past data when training with new data. (c). Architecture-based methods memorize past knowledge by using previous network architecture and learn new knowledge by expanding the network's architecture. The expanded network can be neuron-level or module-level.
• Figure 4: Two different types of data recording for traffic flow
• Figure 5: A crossroads scenario for trajectory prediction. The solid line represents the agent's historical trajectory, and the dashed line represents the predicted trajectory. We observe a green vehicle and pedestrian stationary, while an orange vehicle and pedestrian are in motion. Their trajectories exhibit a degree of predictability. Conversely, the trajectory of the yellow vehicle and pedestrian has just started, introducing significant uncertainty. The objective of trajectory prediction is to anticipate future behavior within such intricate scenarios precisely.

Theorems & Definitions (3)

• DEFINITION 1
• DEFINITION 2
• DEFINITION 3