Ding "ML - Compiling Tradeoffs"

From data-intensive applications that include voice, image, and facial recognition, to using deep neural networks and natural-language processing software to root out hate speech and democratize computing, to developing software that enables machine-learning (ML) models to operate across hardware platforms, UCSB computer-science engineers are at the forefront of this AI-intensive moment

Compiling Tradeoffs

Like any software, machine-learning models run on devices, and devices vary in their available computing resources. That’s why the same ML model that runs on your desktop computer, with its abundant energy and powerful graphics processing unit (GPU), can drain the battery of a smartphone, which has far fewer resources, in less than a half-hour.

UCSB computer science assistant professor Yufei Ding designs the lynchpin software — called a high-performance compiler system for machine learning — that makes it possible to run an ML model on any device “We want to deploy all kinds of ML models efficiently on devices having diverse computing resources,” she says, “and we want both high processing speed and low energy consumption.”

Her compiler system takes into consideration the device provided and the ML models written in a domain-specific language, say, one for bird watching. The compiler automatically optimizes the ML model to run on the device.

"It may change your model a bit, eliminating redundant computation to reduce energy consumption," Ding explains. “So, while, the full model running on a desktop might achieve 95-percent accuracy, it can be made ten times more energy efficient to run on your phone, at the negligible cost of less than 1-percent accuracy degradation.”

To read more about the CS Department's Machine Learning research and affiliated faculty (A. Singh, Y-X Wang, Wm. Wang, Y. Ding, X. Yan):

COE/CLS Convergence magazine (Fall 2019) - "Hot Topics: Computer Science – Machine Learning" (full article pg. 19)