From the course: DJ Patil: Ask Me Anything

How do you identify algorithmic biases?

From the course: DJ Patil: Ask Me Anything

Start my 1-month free trial

How do you identify algorithmic biases?

“

(upbeat music) - How do we identify if these algorithms contain bias? You know, you're supposed to be using a computer program because it's objective. But in reality, it's written by humans. So, how do we deal with that? - This first tenet that I think about is, what does it mean to do data science responsibly? How do you do it with intent? How do you do things with purpose? The beginning of that is just because we can, doesn't mean we should with data. Should we try to figure out advertisements for women who are pregnant? Well in some cases, maybe yes, some cases no. Context matters. But also, how you actually implement this. Also matters. Do we have oversight? Do we have the people who might be impacted by using this data? What happens with them? These are our questions that aren't of only data scientists. These are questions that are far broader. And we have to think about the entire set of implications that are out there. And there's no solid, easy answer here. There's just progress that has to be made across all these fronts of asking, are we okay with this? How do we get external input? All those other pieces have to come together to actually ask, are we comfortable with this? Would the people who're impacted be comfortable with this? And what are the unknown unknowns, the implications that are out there, that we can't see coming. In some cases, the data that's even going in here is terrible. Could be terrible because it's biased. Number two, is we also have to remember in many cases, there are opportunities where the algorithms could do sufficiently better than humans. Because humans are biased also. And then in cases where algorithms can actually take out the bias. There are bail calculators that basically try to assess your risk for bail. There is a set of them that are out there, they're racist. There's also a set of them that actually help flag when there's disparity that is there because of the judge may have a bias already. Saying like, this is an outlier. Maybe we should look at this. And those bail systems actually use data in a very careful way, and are good. The key part there is that you're not handing entire control over to the algorithm. You're using the data and everything to augment somebody, and the algorithm in a way that helps improve their decision-making capability. (upbeat music)

Contents