Building a Scalable, Reproducible Ontology System to Improve Deep Learning Data
summary
Nearmap builds a reproducible ontology system that links pixel labels to formal class definitions, stores example images in a git examples folder instead of inside OWL files, and integrates with the labelling tool via a backend API and CI checks so label instructions and IDs stay synced. The pipeline uses Protege + neo4j + GitHub for editing, import and versioning, implements a custom metadata schema and 16 reusable relationships across five layered ontologies, and a POC finds 8% ontological conflicts, enabling rule embeddings and custom loss functions to reduce violations.
AI-written summary · about
chapters (7)
- 0:00Intro
- 0:14The Problem
- 8:46Scaling the ontology
- 12:54Adding images
- 15:38Class development
- 17:17Versioning
- 23:25Benefits
related talks
20mBeyond Nodes How Ontology Powers Blockchain in the AI Era
04mUsing OpenStreetMap to create training data
30mRealizing End to End Reproducible Machine Learning on Kubernetes
16mRight data at the right time: Tools demo
37mBuilding a Scientific Taxonomy at Scale
28mthe semantic layer your AI needs
38mBuilding robust machine learning systems
23mCognitive Architecture for Process industries