Wednesday, November 10, 2021 - 2:00pm to 3:30pm
Location:In Person and Virtual ET Traffic21 Classroom, Gates Hillman 6501 and Zoom
Speaker:JACK KOSAIAN, Ph.D. Student https://jackkosaian.github.io/
Reliable and Resource-Efficient Learning Systems via Coding-Theory-Inspired Approaches
Neural networks (NNs) are deployed in many settings, ranging from web services to safety critical systems. This has led to the development of various learning systems for deploying NNs. Learning systems must maintain high predictive performance (e.g., accuracy) while also meeting application-level objectives, such as latency constraints. However, meeting such application-level objectives is made challenging by two often-conflicting requirements: (1) Learning systems must operate reliably despite running atop unreliable, failure-prone hardware. (2) Learning systems must use hardware efficiently.
Techniques used to improve reliability often oppose those used to improve resource efficiency. To balance these conflicting goals, many computer systems leverage coding-theoretic tools, such as error-correcting codes and erasure codes. These tools have enabled resource-efficient reliability in storage, communication, and high-performance computing systems.
This thesis explores the use of ideas inspired by coding theory to improve the reliability and resource efficiency of learning systems. We pursue this through three main thrusts:
(1) We show how properties unique to learning systems can be exploited to more-efficiently integrate traditional coding-theoretic tools into learning systems. As an example, we reduce the execution-time overhead of fault-tolerant, safety-critical NN inference on GPUs by up to 5.3x by exploiting trends in NN design and GPU hardware.
(2) We demonstrate that co-designing coding-theoretic tools with learning systems offers new opportunities to extend the reach of these tools beyond their prior limitations. Specifically, we enable resource-efficient fault tolerance in distributed prediction serving systems by using machine learning to overcome a key barrier in prior coding-theoretic tools.
(3) We identify and exploit opportunities for coding-theory-inspired ideas to be used to improve the normal-mode performance of learning systems, when reliability is not a concern. We show that the throughput and GPU utilization of specialized convolutional neural network (CNN) inference can be improved by up to 2.5x through operating over images combined in a coding-theory-inspired manner and making appropriate modifications to the CNN architecture.
This thesis demonstrates the promise of using coding-theory-inspired tools in learning systems and aims to usher adoption of these tools in learning systems, similar to how they have been used in storage and communication systems.
Rashmi Vinayak (Chair)
Ion Stoica (University of California, Berkeley)
Pramod Viswanath (University of Illinois Urbana-Champaign)
In Person and Zoom Participation. See announcement.