More deep learning for the masses, courtesy of Ersatz Labs

Deep learning is all the rage in the machine learning world right now, and we can add another name to the list of companies selling it — Ersatz Labs. The company has actually been around for a while (as part of a consulting firm called Black Cloud BSG), but it officially launched on Wednesday, complete with a new business model that puts deep neural networks inside a GPU-powered appliance.

Dave Sullivan, Ersatz’s founder and CEO, describes the company’s technology as being deep learning for almost anybody. Its pre-built models are available via the aforementioned appliance or as a cloud service (also running on GPUs), and the company has hidden much of the complexity of training them, but has tried not to dumb them down. He says it’s really for people with some prior understanding of machine learning — people who will be comfortable turning the knobs to tune the algorithms to their specific data, and who will be able to recognize when deep learning isn’t the right tool for the job.

“Neural networks tend to learn complicated patterns, whether you want them to or not,” Sullivan explained, and it’s best suited to solving complicated problems.

A couple screenshots of the Ersatz UI, during the data-formatting process.
A couple screenshots of the Ersatz UI, during the data-formatting process.

Which are exactly the type of problems Ersatz is going after. Although users (the company already has about 2,200 of them) can apply its general-purpose models to whatever they choose — it offers four different models best suited for tasks such as object recognition, time-series analysis and feature extraction — Ersatz itself is putting some effort into luring users from the financial services and medical fields. Sullivan thinks deep neural networks could be put to work on medical images or other complex data to point doctors in the right direction of a diagnosis, or to crank up the effectiveness of algorithmic trading techniques.

Like Skymind, an open source deep learning company that launched earlier in June, Ersatz is not the product the academic superstars who have developed its major advances over the past decade or so. It’s not a product of any of their proteges, either. Sullivan is 28 and has a bachelor’s degree, although not in computer science. However, he has been coding since he was 12 and has been working in IT for about four years as the founder Black Cloud BSG, a consulting firm whose profits went to funding his new ideas.

He learned deep learning from reading papers, taking courses and generally consuming the wealth of information available on the web. “All of this information is available online,” Sullivan said. “… If you want to learn about it, it’s there, just read it.” Still, he added, “I’m a reasonably smart person, but I think most of the people writing these papers are a lot smarter than I am.”

The video below, from February, shows Sullivan giving an introduction to deep learning and Ersatz at a meetup in San Francisco.


Sullivan thinks there’s a whole generation of people like him who are educating themselves online in areas such as machine learning and artificial intelligence, and who’ll want some tools to work with that are a lot easier than some of the existing lab-developed tools such as Theano or Pylearn2. Getting stuff to to work in real-world settings for real people can be more difficult than building research tools for experts, he noted. Thus, Ersatz hopes its users will only have to know how to use its products — not how they work — in order to get results.

“The thing that we do different is we make deep learning practical,” Sullivan said.

In a field this new — especially commercially — it’s probably going to take a community of people to make sure everything is done right and everyone who wants access to deep learning tools can get them at the level they need. Sullivan pointed to a paper released by leading researchers in December that highlights a seemingly fatal flaw inherent to deep neural networks — depending on how they’re constructed, two inputs that look identical to the human eye will be classified differently by the machines. He calls that type of research a good, but telling, sign that even the brightest minds in the space are still filling out the details about how deep learning works.

“We’re just getting going … and there’s not going to be any one [idea] that some guy comes up with that’s going to be the be all, end all,” Sullivan said. “… I don’t think there’s going to be a PageRank for neural networks.”