Where Inference Runs Matters: An Inference-Location-Aware Methodology for Embedded ML and TinyML Sensor Systems

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Developing intelligent sensor-based systems becomes methodologically more demanding as inference moves from external infrastructure towards the Edge, Embedded ML and Ti-nyML, where a model with strong offline accuracy can be unviable on its target device through limits of memory, latency, energy or hardware–software integration. Established machine learning life cycles such as CRISP-DM, CRISP-ML(Q) and MLOps remain largely agnostic to where inference runs and give limited attention to problem definition, sen-sor-data generation and validation under real conditions. This article proposes a method-ological framework for sensor-based systems in which inference is brought close to, or on-to, the device. It was derived through a retrospective cross-case analysis of twelve projects spanning systems without machine learning, Cloud ML, Embedded ML and TinyML, and organises development into four macro-areas, nine phases and four checkpoints. The loca-tion of inference is treated as an explicit axis that reweights decisions, checkpoints act as decision gates, and three assessment levels are distinguished: offline evaluation, on-device technical validation and end-to-end validation. Thirteen international experts appraised the framework: across 114 verification questions, mean relevance was 3.55/4 and the over-all S-CVI/Ave was 0.92, with checkpoints rated highest. The results support its content va-lidity rather than field effectiveness, which remains for prospective evaluation.

This publication has 40 references indexed in Scilit: