My research centers around developing innovative methodologies for ambient intelligence by designing intelligent mobile, wearable, and IoT systems. Leveraging advanced data science tools, my work integrates machine learning, deep neural networks, and multimodal signal processing techniques to interpret complex sensor data. My methodological approaches include supervised and self-supervised representation learning, sensor fusion, knowledge-guided model optimization, and lightweight model deployment for resource-constrained embedded devices. By integrating these diverse data science methodologies, my ultimate goal is to enable intelligent, human-centric sensing applications that are both practical and accessible in everyday life.
One of my notable research contributions is to transform the usage of the audio systems on IoT devices for ultrasound sensing. This innovation allows the standard speakers and microphones to perform diverse functions beyond the acoustic domains. For instance, I have designed software-defined air temperature and humidity sensors using only the loudspeakers and microphones on smart speakers such as Amazon Echo. By modeling the ultrasound attenuation and sound speed using signal processing and machine learning, I have enabled the smart speakers to monitor temperature and humidity conditions with comparable accuracy to dedicated sensors, without hardware modifications.
My research also introduces multi-modal sensor fusion solutions to develop super-resolution sensing systems, leveraging the complementary strengths of ubiquitous IoT sensors. My work tackles significant challenges in multi-modal machine learning for IoT, including issues related to data/label scarcity and semantic understanding.