Pareto data framework for resource-efficient edge and IoT machine learning using Minimum Viable Data

This paper introduces the Pareto Data Framework, a resource-aware approach for identifying Minimum Viable Data (MVD) for machine-learning inference. Motivated by industrial sensing deployments, it treats data fidelity as an optimization variable and emphasizes coverage over precision in constrained sensing environments. The framework maps the data–resource trade-off space to identify points where additional fidelity yields diminishing performance gains. By optimizing sample rate, bit depth, and clip length, it enables task-specific sensing configurations. Experiments preserve up to 95% baseline audio performance, identify a 12-bit MVD knee in current sensing, and retain 97.5% and 90.1% baseline accuracy in image classification.