What the lab works on

How should a model combine what a camera sees with what a weather station measures? When should one modality overrule another? AICIP studies principled fusion — late fusion across prediction horizons in FujiView, multi-objective masked image modeling in MEDiC and ExPLoRe, and cross-scale consistency in Cross-Scale MAE — along with the self-supervised pre-training that makes these systems work when labeled data is scarce.

Vision has been AICIP’s home ground since the lab’s founding: from automatic target recognition and smart-camera networks through face analysis (including the widely used UTKFace dataset) to today’s transformer-based recognition, detection in industrial inspection, and generative approaches to inverse problems. Dr. Qi has co-authored two Cambridge University Press textbooks on the subject.

A signature strength of the lab. AICIP’s hyperspectral work — unmixing, super-resolution, anomaly detection — earned the IEEE Geoscience and Remote Sensing Society’s Highest Impact Paper Award and continues today with generative approaches to nonlinear unmixing. Recent work extends to global-scale satellite change detection (IARPA SMART) and self-supervised pre-training tailored to the multi-scale nature of overhead imagery.

The lab applies vision, reinforcement learning, and agent-based AI to care settings: adapting a virtual-reality music intervention to a person with dementia in real time, training socially assistive robots to respond to cognitive and emotional state, and building air-quality alert systems for at-risk homes. This work spans collaborations with UT engineering and nursing partners at Duke, supported by the NIH National Institute on Aging.

If training is a trajectory through weight space, can we predict where it’s going? The lab uses Koopman operator theory to model training dynamics — predicting weights several epochs ahead to accelerate optimization (PDT, ICLR 2026), stabilizing proximal policy optimization (KIPPO, IJCAI 2025), and measuring when and where training is structured enough to be predictable at all.

AICIP’s methods increasingly drive domain science and engineering: the ATHENA testbed applies AI to automated materials discovery as part of a national cloud-laboratory network; USDA-funded collaborations monitor poultry welfare and dairy-herd health with low-cost vision and IoT sensing; and the lab advises UT’s EcoCAR team, continuing a line of radar–camera fusion research for connected and autonomous vehicles.