Evaluating Deep Learning Models with Smartphone Sensor Data
w/ Hyunbin Kim
- Background
- IoT technologies are integrated in smartphones and smartwatches
- Sensors can monitor physical activities → Human Activity Recognition (HAR)
- Priority lies in evaluating sensors
- Aims to explore the performance of accelerometer through deep learning model
- Data: UCI HAR Dataset
- Test subjects: 30 volunteers aged between 19 and 48
- Samsung Galaxy smartphone
- 6 kinds of activities: Walking, Walking Upstairs, Walking Downstairs, Sitting, Standing, Laying
- 50 accelerations of coordinates x, y, z per second
- Pre-processed with noise filters
- 128 readings per window
- Experiment with 5 types of RNN LSTM models
- Different types of layers for each model
- Epochs = 50, batch size = 64 fixed
- Focusing on minimizing fluctuations and gaps between training and validation curves:
- LSTM itself has good performance
- Model with Conv1D layer has best fit for our HAR data
- GRU does not work efficiently with large data