User Avatar

AI Product Leader 🦊 | Author http://Model47.com | Co-Founder #PrimaDonaStudios & #TowardsArcadia | Speaker | Writer🚢

The other day on #RAI LinkedIn, I came across an interesting paper about a concept called "TIBET," which talks about detecting and analyzing biases in text-to-image models. Thought it'd be fun to have a look in #RAI 12/100.

TIBET stands for "Text-to-Image Bias Evaluation Tool," which is a great name!

I think this research is pretty useful to go through, even if you're not an academic (which I'm not & never will be again)!!!

Just the way they're tackling the problems of identifying biases in these models - and offering methods that are comparable over time and across AI providers - is helpful to go through.

They take a prompt and then look for things they call "bias axes." A "bias axis" is a category representing potential biases (like gender, ethnicity, or age), which can influence which images text-to-image models generate.

These bias axes vary, according to what's in the prompt. TIBET helps suggest these, using GPT-3 and counterfactual reasoning.

What is counterfactual reasoning? - you might ask (I did).

It's taking a prompt, then coming up with alternative prompts (counterfactuals) by changing "bias axis" attributes in the original prompt. Then you run all the prompts through the model, and compare the resulting images to identify biases in the model's outputs.

The researchers also come up with quant metrics like Concept Association Score (CAS) and Mean Absolute Deviation (MAD) to quantify these biases.

It's inspiring to see frameworks like this one, helping to enhance fairness in TTI models, integrating with bias mitigation tools, and promoting ethical AI development despite challenges like inherited biases from its components.

Next up (at some point in the future) - figure out how to use it for a real project!