Thought Leaders in Artificial Intelligence: Gurjeet Singh, Co-Founder and Chairman of Ayasdi (Part 4)

Sramana Mitra: By rolling out your software in a risk operation that has such large numbers of manual labor, how many people can you replace with the software?

Gurjeet Singh: At this point, the banks are not talking about replacing people. They’re worried about capturing risk. They are more interested in making sure that they capture every last bit of risk and prove to the regulators that they are doing everything possible to account for anti-money laundering and to stop it in its tracks.

Sramana Mitra: As the founder of the company, what is your assessment? Can this be completely automated?

 At some level, people looking at this stuff cannot be as effective as AI doing it.

Gurjeet Singh: I think it cannot be automated. The reason is not technical. From a risk perspective, you want people looking at this stuff. I’ll give you one example. Imagine that you’re building a credit approval model for a bank. What you would notice is that if you use their prior data, the models will just come out biased because the data itself is biased. It’s not a good thing to carry the bias into the future with you.

At least for the foreseeable future, it’s very important to have human operators in the loop who can actually intervene and understand what the models are doing and what they are saying. We need to be able to have a system that adjusts the biases that are inherent in the data itself.

Sramana Mitra: I’m going to ask you a follow-on question on this. It sounds like you could train your algorithms to detect and correct bias, can you not?

Gurjeet Singh: You can build algorithms to detect bias but you cannot really correct it. In the instance that I’m mentioning, the bias in the data exists because you have a lack of data of a certain class. Imagine in a country like India, you have so many people who are unbanked. The problem is you don’t have enough data of that category. No matter what you do, you can’t correct it. You have to see outcomes from it.

Sramana Mitra: What you’re describing, though, is a corner case. When you see segments of population where there’s bias, you can isolate that and you can put manual resources on those segments. If you look at 100 million accounts in the US, that corner case doesn’t apply to 100 million accounts in the US. That particular bias would not impact that dataset.

Gurjeet Singh: That’s correct.

Sramana Mitra: As long as you control which datasets you’re going to automate and then which datasets require some manual intervention, you can vastly reduce the requirement for manual intervention.

Gurjeet Singh: Absolutely.

This segment is part 4 in the series : Thought Leaders in Artificial Intelligence: Gurjeet Singh, Co-Founder and Chairman of Ayasdi
1 2 3 4 5