With AI technology growing rapidly, it’s essential for the government’s private-sector partners to adopt three core principles for AI intended for government use, the U.S. Patent and Trademark Office’s (USPTO) senior technology official said.
Jamie Holcombe, CIO at USPTO, discussed the importance of those core principles at March 12 event organized by Nextgov/FCW.
“In essence, I’m championing three different governance principles,” Holcombe said. “The first one is data sources. If you don’t understand your data source and have a golden data set in order to train your algorithms, you’ve really screwed up. Because you have to start good – you have to start knowing what your data is.”
“Number two, data algorithms. Data algorithms need to be certified, they need to be understood, they can change, and they need to learn from one another,” he said.
Holcombe said the third principle involves industry taking accountability for the algorithms they produce.
”Number three is the one that I challenge industry to get done, because nobody’s doing it,” he said. “In order for this to legally take hold, people need to be held accountable. In order to have accountability, what you need to have in these algorithms is traceability,” the CIO said.
“What do I mean by traceability? When you run artificial intelligence against large data sets, what happens is you can’t replicate many times a certain run, because it’s already done it, you’re never going to get the same results as you did,” he said. “But if there’s a traceability algorithm, where it certifies or authenticates ‘I went through this algorithm at this time with the date time stamp,’ then you have the traceability for accountability and for legal authority,” he said.
Holcombe also said he thinks that some pervasive AI notions, such as “garbage in and garbage out,” are becoming less of a problem in the fast-paced world of AI technology development.
“I don’t believe that’s true anymore,” he said. “Because with algorithms that learn, you can actually filter out the garbage. So when you put garbage in, if you’re getting garbage out, you have bad algorithms. And in a way, you really need to figure that out,” he said.
