Richa Sachdev, Head of ML Engineering at Vanguard, discusses her approach to ensuring ML models are developed ethically and used responsibly
As organizations get to grips with the practical issues around ensuring AI and ML is used ethically, a lot of effort needs to go into helping business stakeholders understand these technologies. In this week’s Business of Data podcast, Richa Sachdev, Head of Machine Learning Engineering at investment firm Vanguard, shares how she’s ensuring her team puts ethical data at the center of its strategy.
Principles for Ethical Model Development
Sachdev’s team’s primary role includes developing recommendation systems for funds and using data to track customer interactions to support Vanguard’s sales and marketing functions. For Sachdev, doing this ethically means focusing on issues such as privacy, explainability and bias.
“As engineers, we can be proactive about governance by redacting unnecessary information when we’re creating a model,” she says. “Of course, we don’t want to redact everything because the model will lose value. But I don’t need a person’s Social Security number, their religion or their criminal history.”
“We have to ensure that we are not introducing any known or unknown bias in our model baseline,” she continues. “There are a lot of statistical tests that are available in our toolkit for training or testing models. So when we get the outputs, we can compare results to see if something applies to a general population or just a small sample to avoid problems downstream.”
Everyone is Responsible for Using AI Ethically
Sachdev is proud of the strides her organization is making towards data analytics maturity. While there are still departments that don’t understand analytics function, many are making the most of it.
Leveraging analytics cannot be a standalone function, she says. But at the same time, everyone who uses AI within a business has a role to play with respect to ensuring those systems are applied ethically.
“There isn’t a single party that can ensure that everything goes well with ethical data,” Sachdev notes. “Achieving this should be part of the CDAO’s strategy and part of leaders’ key responsibilities. Everything should be connected by a common thread.”
She concludes: “I was in an internal conference, hosted by my department and the data and governance department, where we discussed what ethical AI really is. A lot of deliberate work needs to go into bringing everyone to the party.”
- Consider the ethical implications of each use case. Behaving ethically will often require data scientists to redact unnecessary personally identifiable information (PII) or build explainability into models
- Proactively combat ML bias. Enterprises should develop processes to search for and remediate the many kinds of bias that can lead to unfair model outputs
- Everyone is responsible for using AI responsibly. Stakeholders much be educated on how to get the most out of AI systems and how to do so ethically