Extending SpArSe: Automatic Gesture Recognition Architectures for Embedded Devices

Juan Borrego-Carazo, David Castells-Rufas, Jordi Carrabina, Ernesto Biempica

Research output: Chapter in BookChapterResearchpeer-review

Abstract

Neural Architecture Search (NAS), which allows for automatically developing neural networks, has been mostly devoted to performance on a single metric, usually accuracy. New approaches have added more objectives, such as model size, in order to find networks suitable for resource-constrained platforms. SpArSe [1] is a multi-objective Bayesian optimization framework for automatically developing image classification convolutional neural networks (CNNs) for micro-controller units (MCUs). In this work, we first implement SpArSe and modify it to reduce search time, obtaining similar results regarding accuracy, model size, and maximum working memory but in less optimization time. Moreover, we extend the search space to include recurrent neural networks (RNNs) and add an inference latency objective for time-constrained tasks. Finally, we test our implementation in a gesture recognition task obtaining better results than previous manually tuned approaches for size and performance metrics, which validates the approach and its utility.

Original languageEnglish
Title of host publicationProceedings - 19th IEEE International Conference on Machine Learning and Applications, ICMLA 2020
EditorsM. Arif Wani, Feng Luo, Xiaolin Li, Dejing Dou, Francesco Bonchi
Pages7-12
Number of pages6
ISBN (Electronic)9781728184708
DOIs
Publication statusPublished - Dec 2020

Publication series

NameProceedings - 19th IEEE International Conference on Machine Learning and Applications, ICMLA 2020

Keywords

  • Classifier design and evaluation
  • Face and gesture recognition
  • Neural nets
  • Parameter learning
  • Real-time and embedded systems
  • Wearable AI

Fingerprint

Dive into the research topics of 'Extending SpArSe: Automatic Gesture Recognition Architectures for Embedded Devices'. Together they form a unique fingerprint.

Cite this