Sampling Training Data for Continual Learning in Robot-Cloud Systems

This paper introduces a novel sampling strategy for continual learning between robots and the cloud. This approach addresses the challenge of selecting representative training data from a dynamically generated stream of data by incorporating both uncertainty-based and diversity-based criteria. This balance between exploration and exploitation allows for adaptive adjustment of the sampling rate based on the model's performance, effectively reducing communication cost while preventing catastrophic forgetting. The paper also compares the proposed strategy to existing methods, highlighting its unique strengths in handling dynamic, noisy, and resource-limited data scenarios. Experimental results on real-world datasets demonstrate the superior performance of this approach in terms of accuracy, forgetting, and communication cost. However, the paper acknowledges limitations such as the reliance on labeled data and lack of explicit consideration for privacy and security. Future research will explore addressing these limitations through incorporating weakly supervised or unsupervised learning techniques, encryption, differential privacy, and temporal correlation modeling.