Nervana open-sources its deep-learning software, says it outperforms Facebook, Nvidia tools

Nervana Systems, one of a handful startups focusing on a type of artificial intelligence called deep learning, today is announcing that it has released its Neon deep learning software under an Apache open-source license, allowing anyone to try it out for free.

The startup is pointing to benchmarks a Facebook researcher recently conducted suggesting that the Nervana software outperforms other publicly available deep learning tools, including Nvidia’s cuDNN and Facebook’s own Torch7 libraries.

“We really want to get the tools out there to make it easy for people to apply deep learning to the problem,” Naveen Rao, chief executive and a cofounder of Nervana, told VentureBeat in an interview. “Keeping a closed environment makes it kind of hard for people to try things out and have an idea even for what people can do. If they want the fastest, they’ll come to us.”

The technology ought to catch the attention of companies that have been working with deep learning systems, including Google and Microsoft.

Deep learning involves training systems called artificial neural networks on hefty amounts of information like images and speech snippets and then throwing fresh data at the models to receive inferences in response.

Several startups in this domain have already been acquired, including Madbits (Twitter), DeepMind (Google), and AlchemyAPI (IBM).

Netflix, Pandora, Spotify, and even Snapchat have also been exploring deep learning.

And there are several startups still in this market, including Teradeep, Enlitic, Ersatz Labs, Clarifai, MetaMind, and Skymind.

When VentureBeat first wrote about Nervana a year ago, the startup wasn’t planning to focus on software — specially built deep learning hardware was the top priority. But things have changed; now the startup wants to offer a cloud service drawing on Nervana’s custom hardware that companies will be able to tap for complex computations, Rao said.

The Nervana cloud isn’t available yet, but Neon is. It’s written in Python, a language familiar to data scientists and developers alike. And there’s also a Machine Learning Operations (MOP) Layer, which allows existing deep learning systems like Theano and Caffe to integrate with Nervana’s technology.

“We’ve packaged it up in a way that makes it very easy to use,” Rao said.