Session

DescriptionAdvances in artificial intelligence increasingly depend on the ability to acquire, move, and process data efficiently across the sensing–memory–compute hierarchy. As data volumes grow and energy constraints tighten, the fundamental bottleneck has shifted from arithmetic to data movement, making traditional processor-centric architectures unsustainable for next-generation edge and embedded systems. Emerging paradigms in in-sensor, near-sensor, and near-memory computing offer a transformative alternative by pushing intelligence closer to where data is generated and stored. This special session brings together leading academic and industrial researchers to present a cohesive view of these rapidly evolving directions. The session opens with an industry talk from SK hynix, highlighting the commercial viability of memory-centric AI systems and the rapid maturation of near-memory processing. EPFL contributes advances in compute-memory architectures that tightly integrate arithmetic units within memory banks, yielding orders-of-magnitude gains in DNN inference efficiency. Northeastern University follows with a quantitative framework for evaluating unconventional image sensor architectures, including event-based, coded-exposure, and energy-harvesting imagers, revealing the cross-layer trade-offs that dictate edge performance. Karlsruhe Institute of Technology introduces analog radial-basis-function neural networks for flexible electronics, enabling robust near-sensor inference under strict area and power constraints. Finally, Tsinghua University presents TFT-CMOS hybrid near-sensor compute-in-memory architectures that combine thin-film-transistor arrays and ROM-RAM hybrid CIM designs, demonstrating substantial reductions in data transmission energy and system-level power for edge AI. Together, these talks form a unified exploration of how sensors, analog computing, compute-in-memory, and near-memory architectures can be co-designed to enable the next generation of energy-efficient, autonomous, and scalable edge intelligence.