What a machine learning tool that turns Obama white can (and can't) tell us about AI bias

Abstraction: PULSE face upscaler outputs white faces exposing dataset and structural AI bias

Key points:

• PULSE algorithm (using NVIDIA's StyleGAN) consistently generates Caucasian high-res faces when upscaling pixelated images of Barack Obama, Lucy Liu, and Alexandria Ocasio-Cortez.
• PULSE doesn't enhance original pixels; it generates new faces that, when downsampled, match the input—so there are millions of valid outputs, and the bias shows which are chosen.
• Creators acknowledge the bias is "likely inherited from the dataset StyleGAN was trained on."
• Yann LeCun (Facebook AI) tweeted the problem is data bias; researchers including Timnit Gebru and Deborah Raji argued bias is also structural and social, not just a data fix.
• Deborah Raji: "People of color are not outliers. We're not 'edge cases' authors can just forget."
• Even demographically proportional training data can produce biased outcomes, e.g., a UK-proportional dataset is predominantly white.

Connections: Nvidia · Yann LeCun · Facebook · AI Bias · Generative AI · Computer Vision

Source: https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias