A San Francisco-based startup called Skymind launched on Monday to offer support and services for deeplearning4j, an open source deep learning project it has created. It’s early to tell how much traction deep learning will gain among mainstream companies or even web companies, but the technology does hold a lot of promise. The existence of open source libraries backed by professional services could certainly help spur adoption — especially for a field of data analysis previously relegated to top universities and research labs at companies such as Google, Microsoft, Facebook and Baidu.
Skymind founder Adam Gibson calls what he’s pushing “enterprise distributed deep learning,” and he predicts more companies will be apt to give it a try if it’s packaged in a consumable manner. The deeplearning4j (or DL4j) models are tuned to run easily out of the box, and on standard CPU architectures (although the company might soon offer cloud-based GPU training environments), and are written in Java.
“We just want to make deep learning applicable to everybody else,” he said.
In fact, he added, although it takes lots of brains and computers to really advance the state of art, deep learning environments are surprisingly easy — and flexible — once they’re up and running: “The only thing that’s hard about neural networks is training. Once you have the models, you can run them anywhere.” That includes on laptops or smartphones (see, for example, the DeepBelief SDK for smartphone-based computer vision), or even as JavaScript in a web browser. Skymind is also supporting a Hadoop-based deep learning project called Metronome.
Initially, DL4j includes the same types of deep learning models created and popularized by the researchers now heading up those aforementioned labs — techniques such as restricted Boltzmann machines, convolutional nets, deep belief networks and recursive neural tensor networks — but it’s already working on its own new approaches, as well. If the techniques sound intimidating, perhaps Skymind’s suggested applications will sound more familiar: object recognition, text analysis, voice recognition and even time-series analysis for things like stock-market prediction. (For more on the state of the art in these fields, check out this recent anthology tracking Gigaom’s coverage over the years.)
Skymind will offer its services to users that want support setting up clusters or tuning their models. Once the models have been trained, users should be able to add more data — image, voice or text files — and be confident the models will be able to analyze it, Gibson said, kind of like how Netflix (another company experimenting with deep learning, actually) is able to bring new users into its movie-recommendation system without retraining it constantly.
Although machines still have their limitations, he said, “Deep learning has proven to be the thing that’s able to scale. … [T]he more data you have, the better it gets.” They’re able to extract potentially “billions and billions of observations about one piece of data” automatically, without requiring hand labeling or feature engineering. They require less data transformation than many other approaches to analytics and even machine learning.
He’s planning some new stuff, too, including blending additional mathematical models into deep learning models, and even some new uses cases or capabilities. Theoretically, Gibson explained, models could analyze various different types of data or perhaps handle text and speech simultaneously. “At the end of the day, to a machine, everything’s a number,” he said.
However, Gibson, a college dropout who does teach deep learning at data science training center Zipfian Academy, doesn’t have the academic credentials and name recognition of the field’s well-known researchers — names like Geoff Hinton (Google and University of Toronto), Yann LeCun (Facebook and New York University), Yoshua Bengio (University of Montreal) and Andrew Ng (Baidu and Stanford). His name doesn’t grace dozens of academic papers in the space, as do the other accomplished researchers currently working for large web companies (Li Deng, Quoc V. Le and Marc’Aurelio Ranzato, to name a few), many of whom are products of the big guys’ academic programs.
But Gibson argues that’s not really necessary for what Skymind is trying to do. While those guys are working, at least in part, behind corporate walls on proprietary techniques, and companies such as AlchemyAPI are abstracting the techniques behind APIs (something I’ve argued is actually a really good idea), it’s companies like Skymind (and Ersatz Labs) that are trying to bring these techniques to the mainstream. Kind of like “WordPress for machine learning,” he said, or MongoDB — technologies anyone can use effectively without knowing the nuts and bolts of how they work, but that experts can dig in on and tune to their own purposes.
“If I can show I can reproduce the thing the experts built, that builds brand trust,” Gibson said. “… Then, when I do something new, people will trust it.”
