Invited: A Hardware-Algorithm Vision for In-Sensor Intelligence

In-sensor computing offers a paradigm shift in the design of intelligent edge systems by performing computation as close to the sensor as possible, significantly reducing energy consumption and data transfer overhead. In this paper, we present a short survey and vision of our work on Processing-in-Pixel-in-Memory ($\mathbf{P}^{2} \mathbf{M}$) architectures and dynamic input skipping mechanisms for energy-efficient visual perception. We explore CMOS-based and hybrid emerging memory implementations, including Magnetic Domain Wall MTJs (MDWMTJ), and evaluate our designs across domains like TinyML, neuromorphic vision, AR/VR, and autonomous driving. By integrating pixel-aware masking algorithms into reconfigurable sensor architectures, our systems achieve up to $21 \times$ bandwidth reduction, $53 \%$ sensor energy savings, and state-of-the-art accuracy. We discuss design trade-offs, co-optimization strategies across the stack, and present a forward-looking vision of in-sensor systems that adaptively sense, process, and prioritize data based on task relevance.