Data-Efficient Deep Learning Guidelines for Inertial Sensor Classification

A new research paper addresses the challenge of data dependency in deep learning for inertial sensor-based classification tasks, such as human activity recognition. Given the complexity and cost of collecting large inertial datasets, the study aims to provide data-driven guidelines for determining the optimal minimum sample size required to achieve desired accuracy levels. The researchers developed a unified framework to analyze classification performance across binary and multi-class scenarios, deriving an empirical formula that relates performance to dataset size. Their analysis, conducted on six diverse real-world datasets, consistently showed that accuracy exhibits a logarithmic growth pattern. This finding led to a quantitative stability point metric, indicating that models often reach practical stability with significantly fewer samples than conventional wisdom suggests, enabling more efficient data collection strategies.