Apple’s Carlos Guestrin cautions AI leaders to think very carefully about how they use their data

Carlos Guestrin, Apple’s senior director of machine learning and AI, and University of Washington professor. (GeekWire Photo / Kevin Lisota)

The rise of machine learning has been one of the most exciting developments in modern technology, but according to Carlos Guestrin, one of the oldest principles of computing still applies: garbage in, garbage out.

Guestrin, Apple’s senior director of machine learning and AI, as well as the Amazon Professor of Machine Learning at the University of Washington, reminded attendees at the GeekWire Cloud Tech Summit that all developers of machine-learning technology need to think very hard about the assumptions they make when training their models, and the data they use to train those models, to make sure they reflect a broad set of human experiences.

During one of the most well-received talks of the day, Guestrin highlighted how recent advances in machine-learning technology are being used to solve real problems in society, and discussed how past advances in technology have failed at inclusivity.

“The data that you use is not this abstract thing. It’s something that will define the user experience in a significant way,” he said.

Guestrin, who joined the University of Washington in 2012, founded Seattle startup Turi to bring sophisticated machine-learning techniques to developers who didn’t have the time, energy, or money to learn it on their own. Turi was acquired by Apple in 2016, and Guestrin remains part of the University of Washington’s computer science department.

Professor Guestrin treated attendees to an in-depth lecture, walking through some of the shifts in machine-learning technology since the middle of the 20th century. Lots of fundamental problems standing in the way of artificial intelligence research have been solved, with developments such as specialized chips from companies like Nvidia that can bring immense computational horsepower to these problems as well as the rise of pre-trained models that can let developers use AI without having to learn how to create AI.

Cloud services have also made it possible to access that computing power from anywhere on the planet for far less money than it would take to duplicate that computing power on your own, Guestrin said. And it’s not just on the cloud; he noted Apple’s Vision framework, an image detection algorithm built into iOS that any developer can use in their own applications, as an example of how AI is becoming more accessible.

But as these technologies become more accessible to wider swaths of the population, there are both good things and bad things that can emerge. As more and more women and people of color start to use machine-learning technology, they’ll build applications with far more understanding of the diversity of human life than the white guys who’ve ruled this world to date.

Guestrin explained how the chemicals and processes use to expose early photographs were originally developed with white people as the “perfect” image; early photos of people with other skin tones just didn’t look good because the “ideal” image was designed using white people. That changed over time, but machine learning is still in its early days, and if its practitioners don’t make sure to avoid the same mistakes there could be far worse consequences than an over-exposed photo.

“We have to think about how that data reflects the culture and values that we aspire to,” Guestrin said, leaving those who watched his talk with a lot to think about.

Watch a video of Guestrin’s full talk at the 2018 GeekWire Cloud Tech Summit above.

About the Author: admin

i am as a writer and blogger...

Leave a Reply

Your email address will not be published. Required fields are marked *