Integrated Sensing and Computing Architectures Enabling Smart Vision Sensors

This thesis addresses the critical challenge of the "power and memory wall" in intelligent Internet of Things vision systems. Conventional architectures, which transmit raw data from sensors to cloud processors, are inefficient and create significant bottlenecks in power, latency, and bandwidth. This research demonstrates that by shifting computation to the point of capture, it is possible to create highly efficient, responsive, and secure intelligent sensors. The project explores the co-design of novel hardware architectures, circuits, and algorithms that leverage emerging non-volatile memories, low-precision quantized neural networks, and unconventional computing schemes like the Residue Number System to perform complex AI tasks directly on sensor with minimal energy consumption. The primary contributions of this work are fourfold. First, it introduces efficient event-driven architectures that use on-chip memory for low-power background subtraction. Second, it presents a series of embedded inference architectures that perform analog, in-pixel computation to accelerate various neural networks (BWNNs, TWNNs, and QWNNs) with high throughput. Third, it develops novel privacy-aware mechanisms that use RNS-based encoding and physical integration to protect against sophisticated image reconstruction attacks. Finally, it provides two comprehensive behavior-level modeling frameworks, PiPSim and PINSim, which enable rapid and accurate design space exploration of these complex sensor-centric systems, achieving simulation speedups of over 25,000x compared to HSPICE. Collectively, these contributions provide a cohesive blueprint for the next generation of intelligent, efficient, and secure edge vision systems.

Author dc:creator

• Sepehr Tabrizchi (23291386)

• elect
• Engineering, Electronics and Electrical

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451149.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451149