Google, Microsoft, and Facebook are pioneering a new kind of artificial intelligence.

At Google, it helps drive the voice recognition service that lets you search the web merely by talking into your Android smartphone. At Microsoft, it underpins the new Skype translation tool that lets you instantly communicate with people who speak another language. And at Facebook, a newly assembled team of engineers is exploring how it might be used to recognize faces in online photos. It's called deep learning, and it seeks to remake computing by more closely mimicking the way the human brain processes information, giving machines more power to "learn" as time goes on.

The technology has so much promise, it has sparked a kind of arms race among the giants of tech. Google and Facebook recently hired the two academics who originally laid out the concepts behind deep learning, and earlier this month, Chinese search giant Baidu followed suit when it snapped up another academic at the heart of the movement. But Adam Gibson, an independent software engineer based in San Francisco, doesn't want this new technology locked inside the biggest names on the net. He believes deep learning techniques should be available to any website, company, or developer interested in using them. And that's why he's launching a new startup called Skymind.

"We want to give people machine learning without them having to hire a data scientist," says Gibson, 24, a college dropout who has taught himself the vagaries of deep learning from public academic papers and has served as a kind of machine learning consultant for various companies while teaching courses on the subject through an outfit called the Zipfian Academy.

Alongside another engineer named Josh Patterson, who previously worked for Big Data startup Cloudera, Gibson has built a new library of deep learning software tools that's freely available to anyone, and Skymind will serve not only as a steward for this open source project but as a consultant that will help others use the code to build their own AI-powered online services. Based on academic papers published by some of the deep learning engineers now working for Google and Facebook, the software could help power everything from voice recognition to language translation to the kind of automatic product recommendations you see when you visit Amazon.com.

"We're trying to clone what Google does," says Patterson. And though the project is still in the early stages, Gibson says the code is already capable of bringing deep learning techniques to live web services. "We're handling production-level systems," he says, while declining to name which companies are using it. "At the very least, we're able to reproduce the results that the academic papers are producing."

There are other ways of using deep learning. The academic community that founded the movement offers its own open source software tools written in the Python programming language, and these serve as the basis for Ersatz, a service that lets you tap deep learning algorithms via the internet. But with his open source project, known as Deeplearning4j, Gibson has bigger ambitions. Unlike the academic tools that are already available, his software is built with the Java programming language–thus the "4j"–and that means it can run atop Hadoop, the massive number crunching system that has become a staple inside many of the world's online operations.

Based on software designed at Google, Hadoop is a way of storing and processing enormous amounts of data across hundreds of ordinary computer servers, and this sort of distributed computing power is what deep learning requires. "Hadoop is becoming the system of record for all data," Patterson says. "We need to move deep learning to the data that already lives in Hadoop."

An existing open source project, known as Mahout, already provides a way of running artificial intelligence algorithms atop Hadoop. Overstock.com uses Mahout to drive product recommendations on its popular retail site. But deep learning is something very different from this older breed of AI. According to those who have used it, deep learning comes closer to creating "neural networks" that mirror the way the brain operates. Whereas older AI systems must be "taught" to preform tasks by human engineers in many cases, deep learning algorithms are better at learning and adapting on their own.

David Sullivan, who oversees Ersatz, the online deep-learning service, calls Gibson's project "interesting," and he calls Gibson "a very sharp dude." But he questions whether the move to Java is really that important. "There are more Java programmers out there, but there are probably more machine learning programmers who use Python or other languages," he says.

Gibson and Patterson also argue that Java can eventually provide deep learning calculations at much faster speeds. But Yoshua Bengio, a University of Montreal professor who sits at the heart of the deep learning academic community, says this isn't necessarily the case. "There are other languages which seem better suited for statistical and numerical computation, not just because of the language itself but because of the community around and the set of tools that have been developed around it," he explains.

But Bengio still welcomes Gibson's project–"I'm a big advocate of diversity," he says–and if deep learning is to reach a much wider audience, it must certainly find a place in the world of Java. The language has become one of the primary ways of building big-time web services.

To be sure, the algorithms championed by Gibson are still an awfully long way from cloning the human brain–which means even the artificial intelligence moniker is a big of a stretch–and Skymind is still very much in its infancy. But Google and Microsoft have shown that deep learning can advance the state of the art, and with his startup, Gibson has at least identified the next logical step for this fledgling technology. If he doesn't bring deep learning to the rest of world, someone else will.