Abstract: The rapid emergence of generative AI and vision-language models (VLMs) is transforming multimodal perception and reasoning, but deployment in real-world edge systems remains constrained by bandwidth, memory movement, and energy. I will present a near-sensor software–hardware co-design framework that moves early feature extraction and compression closer to where data are acquired. Instead of treating the sensor, memory, and processor as separate stages, we co-optimize models and hardware to form compact, task-specific representations at the sensor boundary. This reduces off-chip transfer and DRAM energy while preserving downstream accuracy, enabling more scalable edge intelligence for vision and multimodal AI. I will then show how this broader principle extends beyond visual sensing to AI for healthcare. In wearable health AI, efficient inference requires models that can adaptively process multimodal physiological signals, activate sensing only when needed, and remain reliable under incomplete observations. Looking forward, this points toward closed-loop medical systems that not only sense and infer efficiently, but also guide timely actions or interventions. Finally, I will discuss trustworthy clinical AI, where generative models for medical reports must be evaluated not only for fluency, but also for consistency, factual grounding, and verifiable reasoning. Together, these directions outline a path toward efficient, adaptive, and trustworthy AI systems across edge and healthcare applications.
Co-sponsored by: Rice University ECE and Rice IEEE HKN and Student Branch
Speaker(s): Gourav Datta
Agenda:
Presentation at 4:00 to 5:00 pm CDT at Duncan Hall, Room 3076 at Rice University
Room: 3076, Bldg: Duncan Hall, Rice University, 6100 Main Street, Houston, Texas, United States, 77005