Smart Machines

In an oft-quoted interview with Life magazine in 1970, Marvin Minsky, an MIT researcher and pioneer in artificial intelligence, predicted that scientists were about three to eight years away from creating a machine as intelligent as the average human. Such a machine would "be able to read Shakespeare, grease a car, play office politics, tell a joke, have a fight," Minksy said, and it would learn at such a "fantastic speed" that it would reach genius level within just a few months.

Fifty years later, Minksy's vision of a machine on par with the human brain still hasn't been realized - but popular AI tools, such as Google's search engine and Apple's Siri, have become part of everyday life, and machines are learning how to master an array of complex tasks, from operating self-driving cars to spotting tumors to monitoring crops.

"It's no longer, 'is artificial intelligence going to work?' It does work, and there's numerous applications that are out there - medical imaging, credit fraud detection, movie selections - that use sophisticated artificial intelligence algorithms to make business better," says Jeff McFadden, chief technology officer for Xonar Technology, a Largo-based company that designed an Al-enabled security surveillance system to detect weapons. "It's really moved out of that research area and moved into where it's really a commercially viable technology set."

The leaps forward have been made possible by powerful computer processing engines and advances in machine learning techniques. Computers with enough horsepower can crunch large data sets and use a series of algorithms to extract patterns and glean insights from that information. In AI recommendation systems, like that employed by Netflix, the algorithm looks at an individual's viewing history, considers the preferences of other members with similar tastes and evaluates other information to come up with a viewing suggestion. An algorithm in an intelligent credit card fraud detection service, on the other hand, flags suspect purchases by looking for outliers or anomalies that depart from a consumer's normal purchasing behavior.

McFadden's Xonar uses an algorithm to spot concealed weapons based on their ultra-wide band radar signature. The system transmits an electromagnetic pulse toward a person as they walk through it. The pulse bounces back to a receiver. That reflected wave is then analyzed by machine's algorithm to see if it matches the shape, density and other characteristics of various weapons in its learning library. It's more discriminating than traditional technologies, which rely primarily on metal detection, and Xonar can differentiate between a knife or handgun and other harmless metal items, such as keys or money clips, making for fewer "false positives." It's also less obtrusive. McFadden says friends who've breezed through the system in place at the entrances of Ruth Eckerd Hall in Clearwater didn't even realize it was there.

Inspired by neural networks in the brain, these "deep learning" systems can remember and build on observational patterns they find in data. In essence, they become smarter over time, but it's vital to control what data the systems receive. "The big AI in general, deeplearning specifically, learns everything. It learns what you want it to learn, but it learns what you don't want it to learn," McFadden cautions.

In some cases, AI can outperform its human counterparts. A recent study in the journal Nature found that a Google AI system did a better job in predicting breast cancer from mammograms than radiologists did. And a new AI software engineered by the British company DeepMind has created a system of algorithms called AlphaFold that can rapidly and reliably predict the 3-D shape of proteins - a task that usually takes months or years. The breakthrough is expected to speed up and reduce the costs of pharmaceutical development.

Here's a closer look what companies and researchers across Florida are doing with AI technology and why the experts say it will never completely replace humans.