Neuromorphic vision system achieves 193μs inference with 30x latency cut

A team of researchers from multiple institutions (including Jiankai Yin, Zheng Miao, Shuo Gao, Zhong Sun, and others) has introduced a neuromorphic vision system designed for open-world visual intelligence. The system draws inspiration from biological vision and information bottleneck theory, implementing what they call a 'task traction mechanism' on dedicated hardware. Key components include a polarization-sensitive imager and a resistive random-access memory (RRAM) array, which together perform progressive information distillation: light field selection, region of interest extraction, and target anticipation. The entire pipeline completes in just 193 microseconds.

In evaluations across eight challenging open-world scenarios, the system demonstrated substantial improvements over current approaches. For object tracking, accuracy rose by 25.54%; for object segmentation, by 37.73%; and for trajectory prediction, by 36.10%. Perhaps more strikingly, the neuromorphic system achieved an average 30.6-fold reduction in latency compared to state-of-the-art solutions. This combination of speed and accuracy makes it promising for real-time applications like autonomous driving, drone navigation, and robotics—where milliseconds matter. The work is currently available on arXiv (2607.10066) and highlights a path toward energy-efficient, task-oriented vision hardware.