UMass Boosts Deep Learning Research
February 9, 2017
a new cluster of specialized graphics processing units (GPUs) now
installed, the University of Massachusetts Amherst is poised to attract
the nation's next crop of top Ph.D. students and researchers in such
fields as artificial intelligence, computer vision and natural language
processing, says associate professor Erik Learned-Miller of the College
of Information and Computer Sciences (CICS).
Lead researcher Learned-Miller says, "GPUs are critical for modern
computer science research because they have such enormous computational
power. They can address extreme computational needs, solving problems 10
times faster than conventional processors, in days rather than months.
They can run neural network algorithms that are prohibitively slow on
lesser machines. Our new network of 400 GPUs is unusually large for an
UMass Amherst's new GPU cluster, housed at the Massachusetts Green High
Performance Computing Center in Holyoke, is the result of a five-year,
$5 million grant to the campus from Gov. Charlie Baker's administration
and the Massachusetts Technology Collaborative last year. It represents
a one-third match to a $15 million gift supporting data science and
cybersecurity research from the MassMutual Foundation of Springfield.
Deep learning research uses neural network algorithms to make sense of
large data sets. The approach teaches computers through trial and error
to categorize data, much as human brains do. "Deep learning is a
revolutionary approach to some of the hardest problems in machine
reasoning, and is the 'magic under the hood' of many commercial products
and services," says Learned-Miller. "Google Translate, for example,
produced more accurate and natural translations thanks to a novel
Andrew McCallum, professor and founder of the Center for Data Science at
UMass Amherst, says, "This is a transformational expansion of
opportunity and represents a whole new era for the center and our
college. Access to multi-GPU clusters of this scale and speed
strengthens our position as a destination for deep learning research and
sets us apart among universities nationally."
He says the campus currently has research projects that apply deep
learning techniques to computational ecology, face recognition,
graphics, natural language processing and many other areas.
state funds must be used for computing hardware at UMass Amherst, its
Springfield Center for Cybersecurity and for terminals at Mount Holyoke
College and the UMass Center in Boston, the researchers note.
Learned-Miller says he and colleagues are now in the first year of the
grant, during which about $2 million has been spent on two clusters: the
GPU cluster dubbed "Gypsum" and a smaller cluster of traditional CPU
machines dubbed "Swarm II." Gypsum consists of 400 GPUs installed on 100
computer nodes, along with a storage system and a backup system. It is
configured with a leading software package for deploying, monitoring and
managing such clusters.
Not only do the researchers hope the GPUs will accelerate deep learning
research and train a new generation of experts at CICS, but an important
overall goal is to foster collaborations between UMass Amherst and
industry. For example, if MassMutual data scientists design a practical
problem with high computational needs, they can collaborate with
sponsored UMass faculty and graduate students to solve it on the Gypsum