Information & Ethics committee  Hi, I'm Angelina Wang, a graduate researcher in the computer science department at Princeton University. Thank you for inviting me to speak today. I will give a brief overview of the technology behind facial recognition, as well as highlight some of what are, in my view, the mos

Information & Ethics committee  Yes, you are. For brittleness, because we don't really know what the model is picking up on in order to make certain identifications, we don't know what patterns it's relying on. Because humans might know that people are likely to wear makeup and put on glasses, they can control

Information & Ethics committee  Ever since cameras were invented, they have always worked a lot worse on people with darker skin tones. They haven't accounted for different lighting differences. The cameras have always been developed primarily on people with lighter skin tones. A lot of times in different light

Information & Ethics committee  Exactly. The image quality will be different for different people.

Information & Ethics committee  I don't think I have anything else to add to what Dr. Watkins has said.

Information & Ethics committee  Bias amplification refers to a notion of bias that is often thought of as just a correlation in the data. This correlation could be between some particular demographic group and some concept that they are stereotypically related to. Because machine learning models are trying to p

Information & Ethics committee  Sure, in predictive policing, if communities of colour and different neighbourhoods with higher proportions of Black citizens may have higher levels of crime, then predictive policing models may over-report those communities in the future to be more likely to have crime, even if

Information & Ethics committee  I'm not sure what tool you're referring to, but I think measuring these correlations and being aware that even a model with very high accuracy may not be itself amplifying biases and might be creating the same biases that are in the dataset. Even if a model isn't adding additiona

Information & Ethics committee  That's referring to biases in visual datasets.

Information & Ethics committee  I'm sorry. I don't think I'm familiar enough with that.

Information & Ethics committee  I think that because you can acquire facial images without any sort of consent, and that there are so many errors and you don't really know why a model would make a particular decision, then that would go against human rights.

Information & Ethics committee  I'm also not familiar with this.

Information & Ethics committee  Thank you. I think that each model is developed in the context of the different study that it's made by, and so models developed in Asia also have lots of biases. They are just a different set of biases than models that have been developed by Canadians or Americans. For example

Information & Ethics committee  It's very hard to think about, because none of these technologies are ever going to be used in a vacuum, and they're always situated in a particular social context. Even if you had some sort of facial recognition system that worked perfectly, or at least the same across different

Information & Ethics committee  I think that is for me. What the revised tool mostly does is it tries to find different patterns and correlations present in datasets that are likely to propagate into models that are trained on the dataset. It is not guaranteed by any means to find all the possible correlation