University of Twente Student Theses
Exploring the Implementation of Network Architecture Search (NAS) for TinyML Applications
Nieuwenhuis, S. (2022) Exploring the Implementation of Network Architecture Search (NAS) for TinyML Applications.
|
PDF
3MB |
| Abstract: | The use of machine learning (ML) is ever increasing and finding its way in more and more applications, including embedded devices with microcontrollers (MCU). The popularity of inference on MCUs results from the low purchase cost and power requirements of microcontrollers. The deployment of neural networks is challenging since microcontrollers are severely limited by the available memory and storage. This is a problem, especially for the larger and popular convolutional neural networks (CNN). In order to make them fit on a microcontroller, neural networks can be reduced in size by means of quantizing the parameters, reducing precision in the process. Additionally, quantization affects the performance of network inference. These aspects of NN quantization are worth studying to gain knowledge about the trade-off between speed and accuracy. Simultaneously, automated neural network design by means of neural architecture search (NAS) has proven to be able to generate high-performance neural networks that are more efficient than man-made networks. When given proper boundary conditions, NAS-generated networks can be made to fit within the constraints of a microcontroller. In this work, a NAS algorithm targeted specifically to microcontrollers is investigated with quantization in mind. The novelty of this work is the inclusion of quantization levels as a constraint of a search algorithm to construct more complex networks that fit inside the memory requirements of microcontrollers. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 53 electrotechnology, 54 computer science |
| Programme: | Electrical Engineering MSc (60353) |
| Link to this item: | https://purl.utwente.nl/essays/89778 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page