Increasing Machine Learning Training Data to Reduce Selection Bias in AI

Overview

Artificial Intelligence (AI) is growing expeditiously in today’s world and being used for many different real - world applications like healthcare, law enforcement and CCTV, recruiting for companies, etc. This has led to an increase in selection bias, which results in AI developers over representing a specific group of people (usually white and western) and underrepresenting other minorities, so that many facial recognition systems and search engines don’t recognize them or recognize them too much. The purpose of this research is to replicate selection bias and a solution for it by using emotional/type three AI to recognize whether the Teachable Machine software (created by Google) the researcher trained can accurately predict emotions and accurately predict emotions for different skin complexions. The researcher used fair skin toned faces and very small data sets at first and then continue d increasing the data sets used to train the model and made them more representative of other skin tones. The researcher’s hypothesis was supported by the results at the end since the correlation between skin tone, data set amounts, and result accuracy decreased, meaning that selection bias decreased as well.

Competition history

• JSHS 2024 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.