Bias in AI and How to Fix It | Runway

Key Takeaways

I'm sure when you try to generate a photo or a video you probably throw in every description in the book but watch what happens when you leave it all up to the model to decide so first things first what is bias bias is often unconscious tendency to see think or feel about certain things in a certain way biases are somewhat hardwired into our brains to help us navigate the world more efficiently the problem is biases often lead to stereotypes and you think that this is a uniquely human problem but surprise it's not it is a known issue these models tend to default to certain stereotypical representations DT is a a staff research scientist at Runway and she led a critical research effort in understanding and correcting stereotypical biases in generative image models now I think is the best time to fix it because generative content is everywhere we don't want to amplify any existing like social biases there are mainly two ways to approach this problem algorithm and data today we're going to focus on data these models are trained on mountains and mountains of it and because the data comes from us humans here and there our biases start to show up but just like we can uncover and prove our own biases so too can AI models and this process is crucial to ensure fair and Equitable use of AI Technologies the defaults that the model tends to produce are cute towards like younger population very attractive looking women are are men with like really sharp like jawlines one form of beauty that is is pushed onto Us by the society Within These models there are a lot of repetition of certain types of data over indexing and sometimes a general lack of representation altogether we noticed if the profession tends to be like of power like CEO or doctors it does tend to defa to lighter skin tone people and like most likely perceived male as opposed to professions of not very high income do tend to default like darker skin tone and females and this is not a true representation of of the current state of the world this is a big problem we're starting to create solutions for we call it diversity fine-tuning or DFT you might have heard of fine tuning before it's widely used across models to H Styles and Aesthetics the way it works is by putting more emphasis on specific subsets of data that represent the outcomes you're looking for so if you want things to look like anime you would find tune with images like these and this actually works incredibly well even with a very small subset of data the model can learn to generalize from it and this is what diversity fine tuning sets out to do with bias we generated a lot of T proms which are pictures of like female doctor female Doctor Who belongs to a particular ethnicity and used a text image model to generate synthetic images using these prompts PT and our team used 170 different professions and 57 ethnicities and they generated close to 990,000 synthetic images to create a rich and diverse data set to diversity finetuner model it was very exciting to see what we thought was like a simple solution of like augmenting the data and just retraining the model that helped in significantly like fixing the biis diversity fine-tuning is already proving to be an effective way to make text to image models that are safer and more representative of the world we live in I'm being optimistic that we will get to a place where the models are more [Music] inclusive