Using AI to control AI: How to Prevent Creating Biased Datasets

The team responsible for the project scavenged the internet for these images, each one associated with some keywords. However, not all queries (automatic Google searches) were successful. After 100 results, the relevance of images was diminishing. Depending on the source, the team had to set a threshold on the number of collected images. After storing them and analyzing the keywords associated with every image, they were clustered based on common factors, then the labels were chosen by using a statistical methodology.