Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

The University of British Columbia

$59,000.00

Hardware Accelerators for Learning Systems

FY 2023-24

Summary
The desire for increasing levels of automation is driving development and adoption of machine learning. However, the slowing of Moore's Law in recent years is limiting the traditional path to increasing computing capability required to continue this process. While traditionally it would be possible to increase capability by integrating exponentially more transistors in future systems will need to use existing capabilities more efficiently. Of relevance to how this might be achieved is that machine learning can be viewed as consisting of two phases: training and inference. During training large volumes of data are used to learn model parameters. During inference inputs are applied to the model to perform classification or regression tasks. Inference is important in mobile and energy constrained applications. Owing to this, much attention in the past decade has been given to improving efficiency of inference. However, comparatively less research has focused on hardware improvements to increase the efficiency of training. This is important because improvements in machine learning accuracy require increasing computations during training. The objective of the proposed research is to enable order of magnitude increases in training speed for machine learning training in the face of slowdowns in technology scaling. The research program will be divided into two thrusts: Algorithmic improvements and hardware improvements. A key challenge when attempting to speed up machine learning is maintaining the accuracy of the learned model. This challenge exists for both inference and training, however recent work demonstrates that training is substantially more sensitive. This is a consequence of the iterative nature of training: small perturbations can be significantly magnified over many training iterations. Moreover, the end user often would like to see reduced cost but generally not at the expense of significant loss in accuracy. The first thrust of our research will explore approaches at the algorithmic level for reducing total computations during machine learning training. Specifically, we aim to develop techniques to identify and/or encourage computations and intermediate values that may have their precision reduced or eliminated without affecting model accuracy. We will extend these to detecting correlation in computations and data that may exploited for skipping repeated similar computations and/or enabling lossy compression. As improvements in training speed are irrelevant if the resulting model has poor accuracy, we seek techniques that can provide statistical bounds on the expected degradation in accuracy for a given amount of acceleration. The second thrust of our research will involve exploring hardware and/or system software techniques for improving the efficiency of training and inference. We will explore architecture techniques applicable to enabling training within energy constrained environments spanning from datacenter to autonomous robots.
Program
Discovery Grants Program - Individual
Location
British Columbia, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-05411
Name nom
Aamodt, Tor
Department département
Electrical and Computer Engineering
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
59000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1510
Committee name en
Electrical and Computer Engineering
Committee name fr
Génie électrique et informatique
Area of application code
513
Area of application group en
Manufacturing processes and products
Area of application group fr
Procédés et produits industriels
Area of application en
Electrical and electronic machinery and equipment (including computer hardware)
Area of application fr
Machinerie et équipement électriques et électroniques (y compris matériel informatique)
Research subject code
2700
Research subject group en
Information technology
Research subject group fr
Information technology
Research subject en
Technologie de l'information
Research subject fr
Technologie de l'information
Application title
Hardware Accelerators for Learning Systems
Keyword
Computer architecture; Deep neural networks; Hardware accelerators; Machine learning; Machine learning training acceleration; Robotics accelerators
Application summary
The desire for increasing levels of automation is driving development and adoption of machine learning. However, the slowing of Moore's Law in recent years is limiting the traditional path to increasing computing capability required to continue this process. While traditionally it would be possible to increase capability by integrating exponentially more transistors in future systems will need to use existing capabilities more efficiently. Of relevance to how this might be achieved is that machine learning can be viewed as consisting of two phases: training and inference. During training large volumes of data are used to learn model parameters. During inference inputs are applied to the model to perform classification or regression tasks. Inference is important in mobile and energy constrained applications. Owing to this, much attention in the past decade has been given to improving efficiency of inference. However, comparatively less research has focused on hardware improvements to increase the efficiency of training. This is important because improvements in machine learning accuracy require increasing computations during training. The objective of the proposed research is to enable order of magnitude increases in training speed for machine learning training in the face of slowdowns in technology scaling. The research program will be divided into two thrusts: Algorithmic improvements and hardware improvements. A key challenge when attempting to speed up machine learning is maintaining the accuracy of the learned model. This challenge exists for both inference and training, however recent work demonstrates that training is substantially more sensitive. This is a consequence of the iterative nature of training: small perturbations can be significantly magnified over many training iterations. Moreover, the end user often would like to see reduced cost but generally not at the expense of significant loss in accuracy. The first thrust of our research will explore approaches at the algorithmic level for reducing total computations during machine learning training. Specifically, we aim to develop techniques to identify and/or encourage computations and intermediate values that may have their precision reduced or eliminated without affecting model accuracy. We will extend these to detecting correlation in computations and data that may exploited for skipping repeated similar computations and/or enabling lossy compression. As improvements in training speed are irrelevant if the resulting model has poor accuracy, we seek techniques that can provide statistical bounds on the expected degradation in accuracy for a given amount of acceleration. The second thrust of our research will involve exploring hardware and/or system software techniques for improving the efficiency of training and inference. We will explore architecture techniques applicable to enabling training within energy constrained environments spanning from datacenter to autonomous robots.