[Defense] Highly Scalable and Accelerated Kernel Machine Training On Diverse Computing Platforms
Wednesday, April 27, 2022
10:00 am - 11:00 am
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Ruchi
Shah
will
defend
her
proposal
Highly
Scalable
and
Accelerated
Kernel
Machine
Training
On
Diverse
Computing
Platforms
Abstract
The past decade witnessed a massive growth in the volume of data. With the growing popularity of Machine Learning (ML) algorithms, we expect to see around a 40x increase in the size of digital data generated each day by the end of 2025. Synergistic advancements in the processor architecture trends are inching towards exascale computing. This hardware-software complexity makes designing efficient algorithms to train large-scale datasets challenging across various deployment systems. My thesis focuses on bridging this gap between core ML algorithms and diverse computing platforms. I propose highly scalable and performance-centric Kernel Machines, Support Vector Machine (SVM) to perform classification on multi-core CPUs, GPUs, and hybrid multi CPU-GPU systems. To maximize the throughput across diverse platforms, we optimized state-of-the-art algorithms with efficient data distribution techniques, reduced data dimensionality using rank-revealing approximation methods, lower numerical precision, and advanced optimization techniques like Interior Point Method (IPM) for faster convergence rates. Empirical results have demonstrated nearly 1.8x/6.0x/3.5x performance gain for three very large-scale data sets with a minimal trade-off on the optimal accuracy.
Wednesday,
April
27,
2022
10:00AM
-
11:00AM
CT
Online
via
聽
Dr. Panruo Wu, dissertation advisor
Faculty, students and the general public are invited.
