What is it about?
Bias in artifical intelligence systems often comes from biases in the human labelling process of it's training data. We demonstrate how to use the positionality of human data annotators to create more accurate training data in the form of a human-in-the-loop system. This system, Inclusive Protraits (IP) shows statistically significant improvements in all test conditions.
Featured Image
Photo by Clay Banks on Unsplash
Why is it important?
Our system demonstrates the value of improving data quality through considerations of positionality and social theories. Expanding upon this approach can lead to greater gains in data quality, ultimately producing cleaner models, in an easy to implement way.
Perspectives
We hope that this article inspires researchers to look at the way biases enter the training pipeline at the very beginning - with the labelling process. Ultimately, with humans sources of ground truth - we are contending with human sources of bias within the data we then feed to our models. Psychology, and other social sciences, have a long tradition of studying these biases and leveraging these existing insights in the design of labelling tasks can result in higher quality data; empowering both industries and researchers.
Christopher Curtis
Read the Original
This page is a summary of: Inclusive Portraits: Race-Aware Human-in-the-Loop Technology, October 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3617694.3623235.
You can read the full text:
Contributors
The following have contributed to this page
