Files and links (1)
CC BY-SA V4.0, Open Access
Dissertation
An Adaptive Framework for Energy-Efficient Edge AI: from Classification to Synthesis of Images and 3D Shapes
Washington State University
Doctor of Philosophy (PhD), Washington State University
2022
DOI:
https://doi.org/10.7273/000005094
Abstract
A large number of real-time AI applications like robotics, self-driving cars, smart health and augmented (AR) / virtual reality (VR) are enhanced/boosted by deploying deep neural networks (DNN). Currently, computation for most of these applications happens on the cloud due to huge compute, energy and memory requirements. However, moving these applications to the edge like smartphones, AR/VR headsets reduce latency and improves user experience, accessibility and data privacy. Existing solutions utilize high-performance and energy-efficient hardware accelerators lacking customization to software solutions utilizing sparsity and quantization of weights compromising on prediction accuracy. In this thesis, we propose an adaptive framework for energy-efficient edge AI which complements all the previous solutions enhancing the performance on edge devices. We utilize the intuitive thought that easy inputs require simple networks and hard inputs require complex networks. This framework is based on three key ideas. First, we design and train a space of DNNs of increasing complexity (coarse to fine). Second, we perform an input-specific adaptive inference by selecting a DNN of appropriate complexity depending on the hardness of input examples. Third, we execute the selected DNN on the target edge platform using a resource management policy to save energy. We demonstrate the generalization of the proposed solution for three qualitatively different problem settings ranging from convolutional neural networks (CNN) for simple image classification to structured Generative adversarial networks (GAN) for photo-realistic unconditional image generation and Graph convolutional networks (GCN) for 3D shape synthesis. Our experiments on real-world applications on edge platforms in achieving a significant reduction in energy and latency with little to no loss in prediction performance.
Metrics
77 File views/ downloads
70 Record Views
Details
- Title
- An Adaptive Framework for Energy-Efficient Edge AI
- Creators
- Nitthilan Kannappan Jayakodi
- Contributors
- Venkata Janardhan Rao Doppa (Advisor)Partha Pratim Pande (Committee Member)Anantharaman Kalyanaraman (Committee Member)Umit Ogras (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 184
- Identifiers
- 99901019839201842
- Language
- English
- Resource Type
- Dissertation