From Retina to Silicon: The IRIS Framework for Bio-inspired Visual Intelligence
Abstract
Traditional frame-based vision systems are ill-suited for extreme-edge intelligence due to high bandwidth, latency, and power overheads. The IRIS framework overcomes these limits through a unified neuroscience, software, and hardware co-design. Inspired by mammalian retinal circuitry, IRIS embeds multiple key visual functions: object motion sensitivity, looming detection, and motion prediction, directly into the image sensor. Using spatio-temporal filtering and predictive coding, it enables efficient mixed-signal CMOS implementations. Leveraging 3D-integration scheme, IRIS delivers real-time, event-driven feature extraction at lower latency, power, and complexity, enabling sensor-level bio-inspired visual intelligence. Our results highlight that IRIS maintains 98% feature fidelity while integrating three individual visual functions and reducing energy consumption by 2.41×.