课程: Foundations of Responsible AI

What is AI and how does data enable it?

课程: Foundations of Responsible AI

开始一个月试用 购买团队方案

What is AI and how does data enable it?

“

- Media coverage and public discussion about AI is almost impossible to avoid. You may have heard about the IRS aiming to use facial recognition to track tax returns, self-driving car accidents, and tracking software used to monitor students and employees during the shift to virtual school and work during the COVID-19 pandemic. Applications of AI span from self-driving cars to content recommendations on platforms like Netflix. There's so many applications of AI in nearly every industry that it's really hard to keep track. Therefore, we need to make special considerations about who takes the responsibility of the outcomes of AI systems. So what is AI? If you're struggling to define AI for yourself, no need of fret. There isn't an officially agreed upon definition of AI. I like to think of AI as automated decision systems or systems in place that are able to dole out decisions about anything from a rental application to an emergency room triage. We show computers plenty of data, then have them replicate the patterns they identify. What's important to know is that the core of AI is big data. Big data is the basis of AI, because most machine-learning models perform much better when they learn from thousands of examples. That data can be text, images, audio, or rows and columns. But at the core, AI is limited to extrapolating from only what it gets to see. So this means the elements of our world that aren't captured in the data, aren't taken into account. For example, if you look at a healthcare data set with patients and their vital signs, we have a snapshot of one vital sign recording on one day in one person's life. We can't see that they had to call around for a ride to get to the hospital. We don't capture that they're trying to stay afloat with mountains of medical bills. We don't know that a patient was grieving the day that their vitals were recorded. We just know they had high blood pressure and rapid heart rate. But we use their information to train a machine to make predictions about similar people in the future. Maybe on a normal day, our patient wouldn't have had elevated blood pressure, but there's no way for us to tell our machine learning models that. Additionally, machine learning models are parametric, meaning if we've run the exact same model over the exact same data, we can get different predictions. We live in a complex world, and as we build AI tools, we have to critique the data and processes used to train these pattern-recognition machines. While we can easily see the connections between data and flawed machine learning, some issues only arise during the AI design and development lifecycle. In the next video, we'll discover what this development lifecycle looks like and how we can improve it.

内容