A guy reviewing my face recognition model said my training data was mostly white faces, so I rebuilt the whole pipeline

Charlie Roberts · 2026-10-10T03:20:04.037283+00:00

I built a face recognition model last year for a small access control project at my office in Columbus. It worked great in testing, like 97% accuracy, and I was pretty proud of it. Then a friend who does ML at a bigger company looked at my repo and said flat out, "your dataset is basically 80% white guys, of course it scores high." I checked and he was right, I had pulled most of my images from a couple of public sets that skew that way. So I spent about 3 weeks scraping and labeling a more mixed set, around 12,000 images this time with better balance across skin tones and ages. Accuracy dropped to 91% and I panicked for a day, but then I tested on faces it had never seen and the gap between groups basically disappeared. I also added a simple audit script that prints the breakdown every time I retrain, so I can't fool myself again. Has anyone here dealt with a dataset that looked fine until someone pointed out who was missing from it?

🎙️Chatter Central

Loading…

A guy reviewing my face recognition model said my training data was mostly white faces, so I rebuilt the whole pipeline | AI Innovations - Chatter Central