Opportunities 2015

Opportunities 2015: Harness “deep learning” to spot patterns

Artificial intelligence still sounds like science fiction, but one branch of it has made stunning progress

Illustration of a circuit-board brain

(Illustration by Dale Edwin Murray)

Artificial intelligence is one of those concepts that seems to be perpetually around the corner. The results have been fairly unspectacular and confined to university laboratories—until recently. One branch of AI, known as deep learning, has made huge advances in the past couple of years, spawning dozens of startups and promising to bring changes to fields as diverse as medicine and finance, not to mention any company that wants to monitor its brand.

A simple way to think about deep learning is that it’s the replication of how a human brain learns, in a very rudimentary way. Raw data is fed into an artificial neural network, and with enough information and training, it learns to make sense of and categorize the data. The idea of artificial neural networks has been around since the 1950s and experienced a revival in the 1980s, but development was hampered by the fact that it still required a lot of human grunt work. A University of Toronto professor named Geoffrey Hinton has since been credited with discovering a better way to teach neural networks that involves stacking them in layers (hence the term “deep”), with each layer capable of discerning more complex features in the data. Two more developments helped the boom in deep learning: advancements in computing power and the proliferation of data on the Internet. The more information a neural network gets, the better it learns.

The hottest application right now is image recognition. “Photos and video are untapped resources on the web, and there’s a lot of opportunity there to help enterprises and consumers,” says Matthew Zeiler, the Canadian founder of a New York–based startup called Clarifai. Clarifai’s image-recognition technology is powerful enough to instantly identify individual objects in a photograph, discern company logos and determine the demographic characteristics of any people depicted, in addition to a host of other things useful to companies that want to monitor their brands. A beer company, for example, can glean a lot about who is consuming its products and in what contexts by analyzing photos shared on social media, and better focus advertising campaigns as a result.

The medical realm can benefit from deep learning too, especially when it comes to diagnostics. A deep learning algorithm can interpret radiological images to help diagnose cancer more accurately and efficiently. “That’s something we’re really excited about,” says Richard Socher, co-founder of MetaMind, a startup based in Silicon Valley. The company plans to bring its technology to parts of the world where diagnostic expertise is insufficient.

Colorado’s AlchemyAPI, another deep learning firm, aims to build the “next generation Siri, or IBM’s Watson times a hundred,” says founder Elliot Turner. The company has already built a system adept at answering trivia (“even if everybody in our office teams up together, we still lose,” Turner says), but the goal is to provide a virtual assistant people can actually converse with. For example, it could instantly synthesize the findings of thousands of medical papers or help customer service workers deal with consumer complaints. Turner expects to make a product available commercially some time this year.

No one can say for certain how deep learning will evolve. Socher likens the state of the market to the Internet’s early days. “This will spread into most every industry,” he says.