Cross-User Generalization and Personalization for Multimodal Smart-Glass Data

Research assistant project at the People-Centered Computing Lab, USI, advised by Dr. Martin Gjoreski.

• Identified inter-subject domain shift as a key challenge in multimodal smart-glass sensor data, causing significant performance degradation when models generalize to unseen users.
• Pretrained a ViT model with masked autoencoding on multimodal sensor data augmented with Euler angle features.
• Used user identity as a zero-cost supervision signal to study cross-user generalizability.
• Fine-tuned the pretrained model with only 5% labeled data and achieved comparable or superior performance to a fully supervised baseline trained with 100% labeled data across two downstream tasks.