A user told me my facial recognition project was ignoring people with darker skin tones

I was testing a small program I wrote to sort photos by person, and someone on a dev forum pointed out my training set of 1000 images was almost all from a single city in Sweden. I swapped in a more diverse dataset from a public library and the accuracy jumped for everyone. How do you make sure you're not building bias into a project from the very start?

3 comments

3 Comments

rowan7251mo ago

Yeah, the training set thing is so real. I mean, I had a project that was supposed to spot different kinds of birds, but all my pictures were from my backyard. It was terrible at birds from other places. I had to go out of my way to find pictures from other countries and seasons. It's easy to just use the data you have, but you gotta force yourself to look for what's missing right from the beginning.

the_jenny1mo agoMost Upvoted

It's about asking who your tool might fail for before it even runs.

danielr941mo ago

Force yourself to look for what's missing" sounds nice but it's not practical for every small project. Most people just need a working tool, not a perfect global model. Like @rowan725's bird project, sometimes good enough for your own backyard is all you actually need.