Name: This A.I. detects potential bias in other A.I algorithms
Uploaded: 2021-07-13
Duration: 3 min 47 s
Description: Parity A.I. analyzes documentation, identifies risk zones, and recommends methods to assess and mitigate harmful model qualities.

Parity A.I. analyzes documentation, identifies risk zones, and recommends methods to assess and mitigate harmful model qualities.

This A.I. detects potential bias in other A.I algorithms

Tech

Watch: This A.I. detects potential bias in other A.I algorithms

Yeah, parody is a algorithmic auditing platform that is kind of like TurboTax for aI compliance is very complicated process of examining algorithms for bias and discrimination. We actually use a I to address the entire algorithmic systems beyond simply this data inputs that you used to train your machine learning models were able to identify risks of legal action and reputational damage helps companies make sure that they're not harming populations of scale. I founded parody in 2020 and built the technology, I was acquired into twitter where I lead the ml ethics, transparency and accountability team here. I remain an advisor and you know, somebody who's clearly very invested in seeing the company do well. I chose Liz to take over as CEO earlier this year. The idea to create something like parodies patented AI system was built out of years of experience working with companies trying to solve this problem. What I realized was it was not from a lack of will or a lack of capabilities, actually from a lack of being able to identify the unknown unknowns and that's what algorithm tries to do in most cases A I is discriminatory and not because of intense or malice. It's really an unintentional consequence of the technology being relatively new in a lot of ways we're exploring. And just now learning about the different ways that machines accidentally learned discriminatory practices in the field of machine learning fairness, as we call it is really only a few years old, maybe 5 to 10 years now, the field has started to mature enough that we are beginning to see patterns emerge. These things are predictable and preventable. So there are many examples of biases that have been documented in algorithmic system. So for one job hiring algorithms have been known to discriminate by gender. One that amazon was testing out so strongly learned that women should be discriminated against in tech because of the existing biases that is automatically eliminated resumes that it identified as being from people who seem to be women. There are documented examples of biases against women and people of color due to historical biases against these folks. So specifically redlining in the United States has led to algorithms that are that have been trained on historical lending data to discriminate against opposed the predominantly black and hispanic. Um the list goes on and on. But we're seeing real manifestations today in the everyday things that impact our lives. What's interesting about parody is that because we do address this system holistically using model documentation to fully understand the context and the application of the ai. It turns out there's a lot that we can actually automate with regard to checking for this kind of discrimination and harm. We are working down to start helping the world's largest banks and healthcare companies to investigate their models audit them but more importantly, create a repeatable framework that will help them understand these risks, predictably, every time they create a new model or attempt something new with AI. One of the things that I am working on is a VC fund and it's also called parody, it's parody responsible Innovation fund and the ethos behind the fund is that it takes a wide range of companies to enable large enterprise to ensure their technology is being used responsibly. So parity AI is our flagship investment but we also have investments into companies in privacy security, differential privacy, synthetic data home, a market encryption. So we're looking at the wide range of technological capabilities that will be needed to ensure that advanced technologies are being used appropriately.