Computer Science Thesis Oral
— 3:30pm
Location:
Virtual Presentation - ET
-
Remote Access - Zoom
Speaker:
JACK KOSAIAN
,
Ph.D. Candidate, Computer Science Department, Carnegie Mellon University
https://jackkosaian.github.io/
Practical Coding-Theoretic Tools for Machine Learning Systems and by Machine Learning Systems
The use of machine learning (ML) in many domains has led to the development of many ML systems for deploying and training ML models. Beyond achieving high accuracy, ML systems must also use computing infrastructure efficiently and tolerate unreliable infrastructure. Coding-theoretic tools enable many systems to operate both reliably and efficiently. These tools are used in production storage and communication systems, and there is growing interest in their use for distributed computing.
This thesis explores the interplay between ML systems and practical applications of coding-theoretic tools. Specifically, we show how ML systems can be made more reliable and efficient via novel uses of coding-theoretic tools, and how coding-theoretic tools can be expanded in reach and be made more efficient through techniques from ML and systems. We illustrate this via multiple thrusts:
(1) Properties unique to ML systems can be exploited to efficiently integrate coding-theoretic tools into ML systems. First, we reduce the execution-time overhead of fault-tolerant inference on GPUs by exploiting trends in neural network design and GPU hardware. Second, we show how coding-theoretic tools can be coupled with the unique properties of recommendation models to enable low-overhead fault tolerance in training.
(2) Co-designing coding-theoretic tools with ML systems offers new opportunities to extend the reach of these tools. Specifically, we enable resource-efficient fault tolerance in distributed prediction serving systems by using ML to overcome a key barrier in prior coding-theoretic tools.
(3) Ideas inspired by coding theory can be used to improve the performance of ML systems even when reliability is not a concern. We increase the throughput and GPU utilization of specialized convolutional neural network inference by inferring over images in a coding-theory-inspired manner and making small modifications to the model.
(4) Coding-theoretic tools can operate at higher throughput with little developer effort via advancements in ML systems. We exploit similarities between operations in erasure codes, a popular coding-theoretic tool, and those in ML libraries to enable erasure codes to be easily represented via ML libraries, and thus allow erasure-coding libraries to immediately adopt the many optimizations that have gone into ML libraries.
Thesis Committee:
Rashmi Vinayak (Chair)
Phillip Gibbons
J. Zico Kolter
Ion Stoica (University of California, Berkeley)
Pramod Viswanath (Princeton University)
Additional Information
Zoom Participation. See announcement.