Accurately recording what, when, and how much people eat is essential to nutrition research and health interventions. Yet food diaries are burdensome, and no single wearable sensor answers all three questions well. Wrist and thermal sensors can operate continuously and privately but reveal little about food identity or portion size. Cameras provide richer context, but continuous capture and inference increase energy and privacy costs.

This dissertation investigates how a wearable dietary-monitoring system can rely on low-cost sensors most of the time and invoke richer visual models only when useful. I develop a battery-efficient multimodal machine-learning framework that treats dietary assessment as connected tasks rather than one monolithic prediction. The pipeline identifies hand-to-mouth activity from wrist, upward-facing thermal, and camera-supported evidence. A compact thermal trigger determines when to request forward-facing images for eating and drinking recognition, food classification, and caloric-intake estimation. Neural architecture search explores models that balance predictive quality, computation, size, and latency.

The dissertation combines lessons from prior wearable-sensing studies with module-specific methods and evaluations. Low-power candidate discovery can reduce reliance on continuous image processing; triggering separates always-on sensing from higher-cost visual analysis; and task-specific architecture search examines models suited to wearable constraints. Together, these contributions establish a practical design for conditional dietary monitoring. The broader goal is passive dietary assessment that is more accurate and less burdensome while respecting battery and privacy constraints and supporting future on-device and free-living evaluation.