Yottamine Sets Big Data Machine Learning Record


New Software Makes Massive Scale Predictive Analytics Achievable and Affordable

Bellevue, WA September 23, 2013 – Yottamine Analytics, a new machine learning innovator, announced today that it has achieved unprecedented prediction speed, scale, accuracy and price-performance with its Yottamine Predictive Services (YPS) software running on a powerful Amazon Web Services (AWS) compute cluster (CC2.8xlarge). The demonstration represents the largest predictive analytics system ever created on the public cloud.

Yottamine’s software is capable of using large numbers of networked commodity computers to rapidly produce very large, highly accurate predictive models at extremely low cost.
The software’s unique “Core-to-Cloud” hyper-parallel programming enables it to treat many individual computers as a single, enormously powerful calculation engine for building predictive models from big data of all kinds, including database records, web clicks, text, images, audio, video, geo-location data, sensor readings, log files, machine data and more.
In the test, Yottamine ran on 20 networked computers, each with 2 hyper-threaded 8-core Intel Xeon CPUs and 60GB of RAM. Yottamine’s parallel programming reimaged those resources as a single system with 640 processor cores and 1.2TB of RAM.

The control software for the test was the production version of LIBSVM, the most popular machine learning program in the world. The control ran on a dedicated Xeon server, but was able to use only one of its four available processor cores.

Revisiting a classic machine learning problem, the test examined 784 features in 8.1 million example images of handwritten digits to build a classification model that will accurately recognize new cases of written numbers, as might be needed in a forms processing or package sorting system.

The enormous processing required for building the model quickly overwhelmed the single-threaded LIBSVM control software. But highly parallel Yottamine sliced through the Big Data with ease, demonstrating an 1800% speedup over the popular free program.

The control software could not complete the test of a single parameter in three days before being terminated; completion would have taken 40 days. Yottamine processed all parameters in just 6 hours and delivered a human-like 98% accuracy. The software usage price value for building the huge model was just $1,728.

“Yottamine is crazy fast, but it is also highly scalable, accurate, inexpensive, and very easy to use. The technology is transparent to the data scientists and the software eliminates a lot of tedious guesswork for them,” said Yottamine CEO and Founder Dr. David Huang, adding, “That’s the real power of core-to-cloud parallelism – it doesn’t waste compute power and it doesn’t waste people power.”

The Yottamine software runs as an online, on-demand service in a secure, dedicated compute cluster on AWS. It gives the data scientist complete control for building, testing and optionally scoring predictive models. Yottamine users pay only for the compute resources they use, starting at just $3.60 per compute node per hour.

For customers requiring on-premises solutions, Yottamine Analytics will release Yottamine On-Premises (OP) in the first calendar quarter and is seeking beta testers for it now.

About Yottamine Analytics

Yottamine Analytics was founded in 2009 by Dr. Te-Ming (David) Huang to commercialize his original research in machine learning by offering cloud-based predictive modeling solutions that are faster, more scalable, easier to use, and more affordable than other available solutions. Yottamine Analytics now offers a growing family of high performance machine learning software.
Read the gPress Executive Report