Education

Allergy-related reactions such as eye rubbing, palate rubbing, swallowing, and throat clearing occur repeatedly throughout the day but are rarely documented reliably. This thesis aims to develop a machine-learning pipeline that automatically recognizes such events in continuous multimodal wearable-sensor data and subsequently classifies them into distinct reaction types. The proposed approach consists of two stages: Event recognition: Detecting when a potentially relevant event occurs within a continuous sensor stream.Event classification: Determining which type of allergy-related reaction occurred.

Airable is a custom-built smart hair pin or necklace accessory that integrates miniature environmental sensors to support personal air-quality monitoring and pollen-allergy research. The goal of this thesis is to design and evaluate the a basic functional prototype, combining sensing, wireless communication, power management, and a wearable form factor.

This master's thesis investigates deep generative modeling, specifically Stochastic Interpolants and Flow Matching, to synthesize personalized training data for Sensor-based Human Activity Recognition (HAR). While deep learning classifiers achieve high accuracy on benchmark HAR datasets, their performance severely degrades when deployed on unseen users due to inter-subject domain shift. Recent work has shown that Prototypical Networks can successfully address this variability by mapping user-specific movement styles into a continuous latent space using minimal calibration data. This project builds on that success by integrating these established prototypes into a generative framework. The latent representation conditions a Flow Matching model via Classifier-Free Guidance (CFG) during training and enables test-time trajectory steering during inference to generate high-fidelity, subject-specific sensor waveforms. The ultimate goal is to use this personalized synthetic data to train highly accurate, user-adapted downstream classifiers. Candidates need solid deep learning knowledge, strong Python and PyTorch skills, and an interest in generative models and domain adaptation.