Now Available

Thursday, June 30, 2016

Deep Learning

By Kristy McCaffrey

How does your email filter out spam? How does Facebook automatically tag a photo you just uploaded? What powers the crystal-clear voice of Apple’s Siri feature?

The answer is deep learning, a branch of machine learning based on a set of algorithms that attempt to model high-level abstractions in data by using a deep graph with multiple processing layers, composed of multiple linear and non-linear transformations. Said another way—deep learning uses multiple layers of “neural networks” to solve problems.

For example, if a deep learning algorithm was used to evaluate a photograph, the artificial neurons—each consisting of constrained mathematical functions that are designed to recognize patterns—might look for a configuration of pixels that resembles an eye, while another would search for the shape of a lip. These neurons would transfer this information to a higher level of neurons, which would evaluate if the image contained a face. An even higher level would determine if the face was that of a human or some other animal.

The “deep” concept comes from the number of layers; if a higher level determines that the lower levels were wrong, the information is passed back to the lower levels where they learn to better process the data.

Deep learning can be applied to many areas: it can be used on audio files to differentiate between a British accent or a Southern drawl; it can be used to determine whether a review on Amazon is positive or negative; Paypal used it to reduce its fraud rate by 10 percent; the National Oceanic and Atmospheric Administration plans to use it to automate the process of tracking an endangered whale species; medical researchers have developed a tool that might be able to perfectly predict the onset of a seizure in epileptic patients. 

According to Dr. Gregory Piatetsky-Shapiro, an expert on deep learning and machine intelligence, “Deep learning is probably the most important technical development since the invention of the Web in the 1990’s.”

Last year, a computer program utilizing deep learning taught itself to play Space Invaders and 48 other video games with no human intervention. The program was designed to track pixels on the screen and was given the goal of aiming for a high score. Then it was left to simply play on its own. When starting a new game, it would repeatedly lose, but within an hour it had made enough connections between cause and effect to devise a strategy to win.

Although neural networks have been popular since the 1960’s, the technology has recently exploded, partly due to the fact that so many people are willing to share their photos on Facebook, thereby offering researchers the opportunity to test and “train” their tools to improve through practice and learning.

And while many worry that this technology will lead to the artificial intelligence takeover so often depicted in apocalyptic films, most experts believe these fears are unfounded. A human toddler can identify an unknown animal after only a few impressions of a photo, but it takes deep learning machines billions of images to recognize the same. And while that computer program did master Space Invaders, it has yet to conquer the complexity of Pac-Man.

 * * * *

Works Cited
Flaherty, Joseph. “Inside Our Computers’ Brains.” Sky Magazine. April 2016.

No comments:

Post a Comment