Ersatz Thinks Up a GPU-Powered Neural Network
If you want to use neural networks to build models that can learn from data in human-like ways, but are having trouble figuring out where to start, you may want to check out Ersatz, the name of the new GPU-powered deep learning platform officially unveiled today by Ersatz Labs.
The domain of machine learning is advancing very quickly at the moment, driven by humankind’s rather sudden (and entirely insatiable) desire to keep and understand all data. There are also big advances being made within the machine learning sub-discipline of neural networks, so-called because of the way the algorithms mimic how the human brain works.
These advances are on public display in the open source arena. Whenever a researcher at Google or Microsoft or other outfit makes a breakthrough in neural networks–and there have been many of them over the past five or six years–all the gory technical details are typically published on the Internet, enabling anybody with the requisite skills to incorporate them into their own work free of charge.
And therein lies the problem: Building a neural network requires a high level of data science and developer skills. Many of the leaders in the field have Ph.Ds in math, statistics, or computer science. But finding the talent to build a homegrown deep learning platform requires an equally deep bank account these days.
The disconnect between supply and demand is what drove Ersatz Labs co-founder and CEO Dave Sullivan to develop a pre-built neural network that allows customers to leapfrog past the “How do I build it?” part and quickly get to the good stuff, which is “How can I use it?”
“If you are looking to do deep learning, and you’ve tried the open source stuff and found it too complicated, you should really try Ersatz because we’re ready to go and we’re the only guys on the market right now that offer something like this,” Sullivan told Datanami this week. “If I wasn’t building a tool for deep learning, if I’m building an application for deep learning, then I would seriously consider buying something off the shelf rather than building it myself.”
The Ersatz platform has been in beta for the past year-and-a-half, and has been tested by more than 2,000 people across a variety of industries. The platform is available now, either as a pre-built appliance equipped with NVidia GPUs and Intel CPUs, or as a cloud-based offering running on Ersatz’s gear. The appliance starts at about $50,000, while the cloud-based offering starts at $.40 per minute. (Some companies do not want their data in the cloud, Sullivan said.)
Ersatz Web-based interface was designed to mask the complexity of building a neural network. After loading the data, which could be anything from images to text to time series data, the software guides the user through the next steps. “We make that really easy,” Sullivan said. “Press a button, check a couple boxes, and boom! You’re training a neural network on GPUs.”
Depending on the data type, Ersatz will recommend a different algorithm to train the neural network. Convolutional neural networks are good at crunching image data, while time-series and sensor data do best in recurrent neural network models. The platform also has pre-built multi-layer perceptron (MLP) models and autoencoder models available, which are good at extracting features and dimensionality from data.
Actually training a model in Ersatz still requires a data science expertise and experience working with different data sets. This product is not something that your typical CEO or CFO can pick up and be a pro with on day one. Getting the model trained on the appropriate data is the difficult part. After the model has been trained, actually using the model to make predictions is relatively easy; Ersatz exposes a RESTful API for that purpose.
The Ersatz machine has already been applied to a range of problems in a variety of industries, including genomic research, stock market predictions, risk management, and medical imaging. The ability of machines to quickly and accurately identify potentially important features in a medical image, such as an MRI or a mammogram, will be one of the most popular uses of neural networks going forward, Sullivan predicted. Building a model of human behavior will also allow advertisers to predict what advertisements they’re likely to click on (a use case common in the Hadoop world as well).
There are all kinds of applications that can benefit from Ersatz’s capabilities in the areas of feature engineering and feature learning. “Feature engineering helps when you actually know about the problem domain and you can summarize it. It’s really attractive to people particularly with the kind of data sets that we’re now dealing with, where they’re very large and also very complex,” he said. “But when you don’t really know what you’re looking for, or specifically what the signs are…this concept of feature learning really starts to make sense.”
GPUs are about 40 times faster at training neural networks compared to traditional CPUs, Sullivan said. He used NVidia’s CUDA library to optimize the Ersatz neural network software to run on GPUs. Just as GPUs are staking a claim in the HPC world, they’re also finding application in the world of big data and machine learning.
“In 2008 and 2009, some people said ‘Hey let’s try to write some neural network software that uses GPUs and see if that’s any faster. It turns out it was–like a lot faster,” Sullivan said. GPUs “happen to be good at the types of calculations that you need to render 3D scenes for video games, but that turns out to be exactly the same kind of math that you need to train a neural network, that you need to teach it new information.”
Sullivan, who has been programming since age 12, started Ersatz several years ago as a “random harebrained side project” to his day job at a consulting company in the San Francisco Bay Area. “The more I read and learned about it, the more I came to believe that this was going to be the way that machine learning was going to be moving,” he said.
The lack of specialists in the field led Sullivan to his business plan. Ersatz Labs, which gets its name from the German word for “artificial,” may have a lead in a small market now, but Sullivan doesn’t expect the market to stay that way for long, as the interest in neural networks and deep learning is increasing rapidly.
“A lot of people are teaching themselves. I don’t have a Ph.D. I taught myself basically through experimentation,” Sullivan said. “It’s complicated. There’s a lot to learn. But it all makes a lot of sense once you learn it and it’s all pretty intuitive really.”