Siri, Tell Fido To Stop Barking: What's Machine Learning, And What's The Future Of It?
Machine learning is an integral part of Pittsburgh's tech economy, thanks to Carnegie Mellon University's position as one of the nation's foremost research centers on the topic. That's enticed tech giants such as Google and Uber to set up shop in the Steel City.
Pittsburghers have varied knowledge on what machine learning is.
On a crisp afternoon on Carnegie Mellon University's campus, Adeline Mercier of Squirrel Hill was walking with her young daughter on campus. She said her husband works in machine learning.
"It's when you train a computer, for example, or a program to learn something," Mercier said. "To optimize something or to automize something."
Alex Xu of Oakland was a little more specific.
"It's like applied statistics used for understanding how patterns can get recognized," Xu said.
A handful of people, including CMU student Karen Abruzzo, were unfamiliar.
"I've heard of it, but I don't really know what it is," said Abruzzo.
Tom Mitchell, a professor of machine learning at Carnegie Mellon University, said machine learning aims to answer the question of how to make computers improve automatically with experience.
"Humans are the best learners, much better than machines are these days, but the idea is similar," Mitchell said. "For example, if you learn to play chess you start out knowing the rules but not the strategy, and you make mistakes and learn from those and become better."
Mitchell said computers learn similarly. For example, he said people know how the recognize their mother in a photograph but can't write down an algorithm for how. An algorithm is like a recipe for a computer, telling a computer step by step what to do.
"Today it's very easy to train a computer program to recognize your mother by showing it photographs, you say in this photograph, here's my mother, this photograph, my mother is not in this one," Mitchell said. "If you give it enough of those training examples, machine learning algorithms look at the details and they find what's common to the positive examples that distinguish it from the negative."
Mitchell said one of the earliest commerical uses of machine learning was credit card fraud detection, trained on hundreds of millions of examples legitimate and illegitimate transactions. This system is still used today.
Mitchell said machine learning is also used to diagnose skin cancer in a blemish.
"Computers are now at least as accurate as carefully trained doctors," Mitchell said. "It's simply because it can examine more training data than people will see in a career."
In the future, Mitchell predicts machine learning will apply to more and more parts of life. He said it will likely become more similar to how people learn, too, especially when it comes to systems such as Siri and Alexa.
"I think in the future, you'll be able to use [smart devices] by saying, that sound you just heard was my dog barking, and whenever my dog barks and you don't hear me respond, I want you to say in my voice, 'It's okay Fido, calm down,'" Mitchell said. "I think in the coming decade we'll be able to teach them the same way you would teach me to do something if I was your assistant."