Kate Park: Data Engines for Vision and Language

Data Quality

The importance of high-quality data in machine learning cannot be overstated. Kate Park emphasizes that data scaling experiments often reveal performance improvements, even if they plateau over time 1. She explains that the data engine approach allows for consistent performance enhancements, particularly when architectural changes are not feasible 2. This method was pivotal in developing features like Tesla's Navigate on Autopilot, where data-driven improvements unblocked feature releases 2.

The proof is in the pudding. And I increased my confidence in this approach every time a release was unblocked by the data engine.

--- Kate Park

The data engine's success underscores the critical role of data quality in achieving machine learning breakthroughs.

   

Data Labeling

Data labeling is a crucial yet challenging aspect of machine learning, involving both human oversight and AI assistance. Kate Park explains that for computer vision, labeling involves segmenting images and adding bounding boxes, while for large language models, it can be as simple as creating prompts and responses 3. She highlights the evolution of data labeling in projects like InstructGPT, where human labelers initially provided supervised fine-tuning data and later engaged in preference ranking for reinforcement learning 4.

A human labeler that we recruit would write both the question and the ideal response that the model should mimic.

--- Kate Park

This shift in data labeling practices reflects the growing complexity and sophistication of AI training processes.