Timezone: »

Studying Bias in GANs through the Lens of Race
Vongani Maluleke · Neerja Thakkar · Tim Brooks · Ethan Weber · Trevor Darrell · Alexei Efros · Angjoo Kanazawa · Devin Guillory

In this work, we study how the performance of generative image models are impacted by the racial composition of their training datasets. By examining and controlling the racial distributions in various training datasets, we are able to observe the impacts of different training distributions on generated image quality and the racial distributions of the generated images. Our results show that the racial compositions of generated images successfully preserve that of the training data. However, we observe that truncation, a technique used to generate higher quality images, exacerbates racial imbalances in the data.

Author Information

Vongani Maluleke (University of California, Berkeley)
Neerja Thakkar (University of California, Berkeley)
Tim Brooks (UC Berkeley)
Ethan Weber (UC Berkeley)
Trevor Darrell (UC Berkeley)
Alexei Efros (UC Berkeley)
Angjoo Kanazawa (UC Berkeley)
Devin Guillory (UC Berkeley)

More from the Same Authors