Petuum Platform Promises To Run Machine Learning Programs Efficiently On Any Hardware

Carnegie Mellon Spinoff Helps Users Tackle Ever-Larger Problems

Printer-friendly version

Computer scientists at Carnegie Mellon University have spent years inventing and perfecting a platform that uses workstations, distributed computers, mobile devices or embedded devices to solve large machine learning problems efficiently and effectively, and have now spun off a company, Petuum Inc., to make those capabilities available commercially.

Eric Xing, a professor in the Machine Learning Department and founder and CEO of the company, said the company has already obtained $15 million in initial venture capital funding and expects to have its first products on the market early next year.

Machine learning (ML) and artificial intelligence (AI) technologies are key to innovations such as self-driving cars, speech recognition, computer vision, natural language processing and analysis of electronic medical records, and many other enterprise big data analysis applications.

"In 10 to 20 years, AI and ML will be the dominant workload of any computing device," Xing said. "We need to optimize how AI/ML programs are designed, programmed and run on such devices, especially as these programs grow in size and sophistication. In many areas, such as self-driving cars, current limits on most AI/ML solutions – which are often hand-crafted black boxes – have become a bottleneck."

 "Petuum promises to be a transformative platform, enabling AI/ML programs to be built easily and to be mounted and run on different hardware platforms, using standardized methods that are transparent and repeatable. The technology in the Petuum platform allows the programs to run correctly, quickly, at scale, and using minimal computing resources," Xing said. The company’s vision is for the platform to eventually run on any type of hardware.

"Our platform will make disparate computing devices, from data centers to mobile and embedded platforms, look and function like a single computer," he said. The massive data sets required for many large AI/ML problems already exist across these devices, he noted. Petuum will allow the AI/ML programs to also operate seamlessly across these distributed computing devices. 

Petuum seeks to enhance and expand the use of artificial intelligence and machine learning at the much larger scales possible with distributed computing. Communication between computing devices can be tricky for AI/ML, Xing acknowledged, but he and his Sailing Lab team with collaborators have developed parameter servers, managed communication, and load-balancing methods over the last 3-4 years that automatically keep the devices running in synchrony.

Though other groups have solved machine learning problems using distributed devices, Xing and his team have shown their approach provides an optimal, efficient solution for all types of ML problems, not just certain subsets such as deep learning. The platform thus can support a wide range of applications, such as natural language processing, image and video understanding, and anomaly detection in transaction data.

"We reached a point where we couldn’t go further without capital investment," Xing said, prompting the launch of the company. He expects to hire 30 to 50 people in the next six months and, because of the need for highly trained computer scientists and engineers, intends to keep the company in proximity to Carnegie Mellon University. "Our goal is to build in Pittsburgh, recognizing the strengths of the city and of CMU in helping us obtain the top talent we need," he added.

Xing, who has appointments to the Language Technologies Institute and Computational Biology Department, was until recently the director of the Center for Machine Learning and Health at CMU, which is part of the Pittsburgh Health Data Alliance, a collaboration between Carnegie Mellon, the University of Pittsburgh and UPMC.

Though the launch of Petuum caused him to step down from that post, one of the first products Petuum plans to reveal is a method for assessing disease risk and predicting readmission rates for patients by analyzing electronic medical records and searching for patients with similar conditions. This product could provide new solutions for precision medicine and decision-making problems in the healthcare industry.             

Byron Spice | 412-268-9068 | bspice@cs.cmu.edu