From Tacit to Explicit Knowledge in Machine Learning with a Focus on Computer Vision

Abstract

This paper surveys machine learning, with a focus on computer vision, for operationalizing tacit know-how and skills from observational data in domains such as industrial process compliance and safety, sports motion assessment, and skill-level evaluation. The workflow is organized as a chain from observation, through encoded intermediate representations such as temporal structure and state descriptors, to external outputs. Broad explicit outputs, including scores, flags, and temporal segments, are contrasted with narrow explicit knowledge in the form of verifiable narratives grounded in evidence and linked to evaluation criteria. Because accuracy alone does not ensure operational acceptability, the core challenge is framed as a measurability-acceptability gap shaped by requirements beyond accuracy, notably traceability, verifiability, and accountability. Datasets and methods are reviewed by output format, and generative-AI explanations and suggestions are discussed under constraints that preserve grounding and traceability. Open issues are grouped into reference information, governance, and deployment robustness for high-risk settings.