Amazon reworked Generative adversarial networks to improve clothing searches to better match product descriptions. Generative adversarial networks were initially created in 2014 to produce synthetic images, sing two networks. One network produces fake images, and one that attempts to find fakes, and these two networks compete against each other. However, Amazon was able to give the system the ability to keep old features and adding new ones simultaneously. Amazon was also able to prioritize images with color features more similar to the description.
The new system was able to increase item classification by approximately 22%, the color was improved by 100%, and gender was up by 22%. Overall the new findings could eventually be integrated into Amazon's e-commerce search engine.
Image Credit: Shutterstock