CAGO-ECIL: Cloud-Assisted Genetic Optimization for Edge-Class Incremental Learning with training acceleration

The integration of edge computing and deep learning has significantly advanced edge intelligence. However, implementing incremental learning directly on resource-constrained edge devices remains challenging. Most existing approaches rely on cloud-based training, leading to slow model updates and difficulties in meeting rapidly changing demands, such as in robotics and autonomous driving. To address this, we propose CAGO-ECIL, a Cloud-Assisted Genetic Optimization for Edge-Class Incremental Learning approach. CAGO-ECIL accelerates learning by formulating a learning optimization problem based on quantitative efficiency metrics and using a cloud-assisted genetic algorithm to determine the optimal ratio of new to old samples. This guides edge-based incremental learning to adapt more swiftly while maintaining high performance. Experimental results show that CAGO-ECIL improves accuracy by at least 4.66% and reduces training epoch time by up to 90% compared to state-of-the-art methods. It also achieves competitive average accuracy and average forgetting measures relative to cutting-edge approaches. With a convergence analysis, CAGO-ECIL effectively addresses the unique challenges of incremental learning in edge intelligence.