Development of a Distributed Deep Learning Architecture to Train a Model for Classifying Human Activities Using Tensorflow Lite Running Locally on One or More bangle.JS Smartwatches
Introduction
Bangle.js is a JavaScript-based, individually programmable smartwatch. Furthermore, the device is able to run the Deep Learning framework TensorFlow Lite. Based on IMU-data collected by the smartwatch we want to train a local Deep Learning model which is able to classify human activities performed by the wearer. Models locally trained on the smartwatch(es), shall then interact with a globally trained model using ensemble or Federated Learning practices.
We expect this master thesis to be highly challenging, yet also bearing a high scientific potential. We therefore expect to follow-up the thesis with a scientific publication.
Tasks
- Development and implementation of JavaScript-based network architecture using the Bangle.js
- Locally train a Human Activity Recognition DL model on one or more Bangle.js
- (Optional) Align locally trained model with global meta-model (federated/ ensemble learning)
- Evaluation using state-of-the-art metrics (accuracy, precision, recall, F1-score, LOO cross-validation)
- Evaluation of reliability and energy efficiency
- Preparation of a written report in German or English
- Motivated and independent working
Requirements
- JavaScript (good – very good)
- Machine/ Deep Learning & TensorFlow (good – very good)
- Signal processing & sensors (good)
Working Materials
- Bangle.js website (https://banglejs.com) & TensorFlow website (https://www.tensorflow.org/lite)
- Guest talk Gordon Williams on Bangle.js at University of Siegen (https://youtu.be/73aBUkLissM)
- Energy-efficient Human Activity Recognition using Adaptive CNNs (https://arxiv.org/abs/2102.01875)
Contact
Alexander Hölzemann ( alexander.hoelzemann@uni-sie.. )
Marius Bock ( marius.bock@uni-sie.. )