Poster Abstract: Investigating Fusion-Based Deep Learning Architectures for Smoking Puff Detection

Abstract

Supervised deep learning methods have the ability to extract useful features from raw data when a sufficient volume of labeled data is available for training. However, in emerging application areas such as mobile health, the high cost of data collection often precludes collecting large-scale labeled data sets. As a result, machine learning pipelines based on hand-engineered features remain common. In this paper, we investigate architectures for combining hand-engineered features with deep learning-based feature extraction from raw data to enhance prediction performance on small labeled data sets. We use smoking puff detection from wearable sensor data as an example application domain.