UC San Diego

This study aims to enable deep learning models on resource constrained devices considering system and application requirements. The proven success of deep learning models can be extended into people’s life even more by using them on mobile systems since mobile systems are ubiquitous in people’s lives. However, mobile systems are resource constrained by their nature. The deep learning models require high computation power and resource constrained devices do not have high computation power. This contradiction makes the research for enabling deep learning models on resource constrained devices necessary.

In the first part of the study, we focus on the contention in a system and how it affects our deep learning-based application. Since today’s systems are very complex and handle many tasks at the same time, the tasks can create contention in different parts of the system. This contention may create undesired and unpredictable effects on the applications. In this study, we study the contention in the systems and create a model selection framework which makes deep learning-based applications contention agnostic with limited accuracy cost.

In the second part of the study, we focus on dynamic neural network architectures. The dynamic neural networks are very useful for mobile systems as their conditions and requirements may change frequently. In this study, we develop slimmable early exit deep learning models and an efficient branch selection algorithm. Our dynamic neural network architecture is shown to be effective to adapt changing latency and power requirements.

In the third and last part of the study, we focus on scalable design space exploration to find pareto optimal design configurations for deep learning models on embedded systems. Finding the correct configurations for executing deep learning models on embedded systems is a challenging problem since testing a configuration is costly in time and the design space is vast. In this study, we develop a fast and scalable exploration algorithm that works well for extremely large search spaces.