In a project that Andrew Ng was involved in, teams worked to create solutions via improvements to models and improvements to data to improve detection of defects in steel.  Model-centric efforts resutled in + 0% change.  Data-centric efforts resulted in +16.9% change.

University of Michigan - Ann Arbor

According to Andrew Ng, the left side of this chart represents a traditional view in the machine learning field, which has been reinforced though events like competitions where teams focus their work on making changes to models to improve performance.  Ng posits that this emphasis has resulted in an overfocus on making adjustments to models, when improvements to data quality could lead to better results.

Andrew Ng's Comparison Chart for Data-centric and Model-centric Approaches

DeepLearningAI. (2021, March 24). A Chat with Andrew on MLOps: From Model-centric to Data-centric AI. YouTube. https://www.youtube.com/watch?v=06-AZXmwHjo

A Chat with Andrew on MLOps: From Model-centric to Data-centric AI

Comparison of Improvements to Code and to Data

One way to improve data quality is to focus efforts on assigning labels as consistently as possible.  Finding the best labels for for a few hundred out of 10,000 training set examples can result in great gains in performance.  Ng’s chart on the bottom right shows that changes to decrease noise in a small-to-medium dataset can result in increases in performance that bring it up to par with performance of very large datasets (which are less affected by noise).

Learn Before

Related