Certified Streaming Updates for Machine Learning Models under Limited Throughput.
TL;DR
This paper introduces Scalable Streaming Learning via online Sampling (S3), an efficient model-updating framework designed to handle limited-throughput streaming data while theoretically guaranteeing an approximation to the complete retrained model.
Abstract
The rapid growth of real-time data acquisition creates a critical demand for streaming machine learning at the edge to enhance application performance; however, the sheer magnitude of incoming data often overwhelms the modest computational resources of edge hardware. Consider autonomous vehicle condition recognition, where sensors generate high frequency signals at sub-second intervals, yet edge computing resources can process only a fraction of the incoming data stream. This raises the fundamental question of how to balance the trade-off between throughput constraints, update efficiency, and model utility. In this paper, we introduce Scalable Streaming Learning via online Sampling (S3), an efficient model-updating framework designed to handle limited-throughput streaming data while theoretically guaranteeing an approximation to the complete retrained model. S3 utilizes online sampling to construct throughput-admissible subsets from the continuous data stream, performing Newton-step updates on these representative samples. Through rigorous theoretical analysis, we demonstrate that S3 provides certified streaming updates under both uniform and leverage-based sampling strategies. Experiments on real-world datasets, including vehicle condition recognition, validate our theoretical guarantees and demonstrate practical effectiveness.