Sub-picojoule-per-bit volitional neuromorphic devices for precise targeting and tracking

The technological revolution driven by artificial intelligence has significantly improved the hardware performance, but energy consumption remains a critical bottleneck. The state-of-the-art retinomorphic devices, as core components of artificial intelligence hardware, excel in feature extraction but are constrained by passive attention mechanisms that lack flexibility of actively extracting additional features. Inspired by the human visual system, this work introduces a volitional neuromorphic device with active volitional attention regulation. By leveraging gate-voltage-tunable photoconductance to generate adjustable differential spectral response and employing neural networks to evaluate spectral reconstruction accuracy, the device achieves selective task perception. Experimental results demonstrate a data compression ratio of 1.17% and an extreme information energy efficiency of 0.625 pJ/bit. This advancement not only advances retinomorphic hardware design but also provides a sustainable pathway for energy-efficient hyperspectral imaging and next-generation neuromorphic computing systems. Energy consumption is a challenge in AI hardware. The authors introduce a volitional neuromorphic device that mimics human visual attention, achieving high energy efficiency and precise object tracking with minimal data redundancy.