Deep Learning for (Some of) the People

Deep learning: It’s not just for incredibly well-paid supergeniuses anymore. At least, that’s what startups Skymind and Ersatz Labs would have us believe.

Silicon Valley loves to obsess over things, and this year one of its main obsessions has been deep learning, a catchall term for a kind of artificial intelligence software that enables computers to interpret data in new and sometimes fantastic ways. In basic terms, computers can be given a task and some guidance and then start learning on their own to develop novel solutions to complex problems.

Deep learning techniques have driven some of the latest breakthroughs at Google, Facebook, and Microsoft in areas such as facial recognition and instantaneous translation. In January, Google acquired the deep learning specialist DeepMind for about $500 million, and its rivals have also embarked on hiring binges in the field. According to Microsoft’s research chief Peter Lee, a deep learning whiz can earn as much as a top NFL quarterback right now.

Not every company can afford a Peyton Manning or a Tony Romo. (Yeah, I said it). Heck, most companies can’t even afford an out-of-work swimwear model like Matt Leinart. The result is that a lot of companies want to begin experimenting with deep learning technology but don’t have the means or expertise to do so.

One solution to this problem may have arrived this week, courtesy of Adam Gibson, a 24-year-old, self-taught deep learning expert. Gibson has set up a website called deeplearning4j with a collection of information about the technology. More important, he has created a number of open-source tools that any software developer can download. His open-source code provides the basic framework to grab a pile of data and begin using his algorithms to analyze it.

Gibson has also set up a company called Skymind to complement the open-source site. This is something of a consultancy in that Gibson will help companies assess how deep learning might be of use to them and then guide them in using the software tools to solve their problems. Gibson’s big pitch here is that his technology works well with Hadoop, the popular data analysis software that’s already in use at thousands of companies. “This technology can save companies time and money,” Gibson said. “I want to give people the ability to do this type of data analysis themselves.”

Ersatz Labs has a more formal pitch and will shift on June 11 from testing its service privately to offering it to the public. The startup lets companies upload their data to its cloud service, where they can then put a variety of deep learning tools to work on the information. Companies can also buy specialized computers from Ersatz that have its software pre-loaded on them.

While the startups bill this stuff as deep learning made easy, there’s still a significant amount of geekery required to have any idea what they’re talking about. Ersatz, for example, lets customers pick between Deep Feedforward, Convolutional, and Recurrent networks when configuring their deep learning systems. The Deep Feedforward Net, says the company, “is the one that started it all—the humble but amazingly effective multi-layer perceptron (MLP). These networks are trained using gradient descent and regularized with dropout (among other techniques). This is probably the network to start with.” Got that, Matt Leinart?