Thread by @v_vashishta

Supervised deep learning is limited by label quality. An ontology must be built before labeling begins. That's a graph defining concepts and their connections. Ontologies guide labeling to ensure consistency and completeness.
1/5
#DataScience #MachineLearning #DeepLearning

Any problem space including people introduces multiple, often conflicting ontologies. Datasets ideally have multiple labels and require multiple models to be trained.
2/5
#DataScience #MachineLearning #DeepLearning

Most projects have a single, majority consensus labeling methodology. Where ontologies diverge from or conflict with it, inference will be inaccurate no matter how incredible the models we use become.
3/5
#DataScience #MachineLearning #DeepLearning

This is 1 of many issues that define an upper bound on the complexity of systems we can model reliably. Deep learning is an amazing tool but we need to spend more time teaching and researching its limitations.
4/5
#DataScience #MachineLearning #DeepLearning

We've tried overcoming the complexity bounds with larger datasets and new models. Those lines have reached their limits so it's important for us to evaluate why our approaches fail instead of finding new ones that fail less.
5/5
#DataScience #MachineLearning #DeepLearning