Chen Jiangang, Wen Zhixing, Yang Fan, Bian Renji, Zhang Qirui, Pan Er, Zeng Yuelei, Luo Xiao, Liu Qing, Deng Liang-Jian, Liu Fucai
School of Optoelectronic Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China.
Yangtze Delta Region Institute (Huzhou), University of Electronic Science and Technology of China, Huzhou, China.
Nat Commun. 2025 Jan 15;16(1):702. doi: 10.1038/s41467-024-55701-0.
Neural reuse can drive organisms to generalize knowledge across various tasks during learning. However, existing devices mostly focus on architectures rather than network functions, lacking the mimic capabilities of neural reuse. Here, we demonstrate a rational device designed based on ferroionic CuInPS, to accomplish the neural reuse function, enabled by dynamic allocation of the ferro-ionic phase. It allows for dynamic refresh and collaborative work between volatile and non-volatile modes to support the entire neural reuse process. Notably, ferroelectric polarization can remain consistent even after undergoing the refresh process, providing a foundation for the shared functionality across multiple tasks. By implementing neural reuse, the classification accuracy of neuromorphic hardware can improve by 17%, while the consumption is reduced by 40%; in multi-task scenarios, its training speed is accelerated by 2200%, while its generalization ability is enhanced by 21%. Our results are promising towards building refreshable hardware platforms based on ferroelectric-ionic combination capable of accommodating more efficient algorithms and architectures.
神经重用能够驱使生物体在学习过程中跨各种任务泛化知识。然而，现有的器件大多侧重于架构而非网络功能，缺乏神经重用的模拟能力。在此，我们展示了一种基于铁离子铜铟磷硫化物设计的合理器件，以实现神经重用功能，该功能由铁离子相的动态分配实现。它允许在易失性和非易失性模式之间进行动态刷新和协同工作，以支持整个神经重用过程。值得注意的是，即使经过刷新过程，铁电极化仍可保持一致，为跨多个任务的共享功能提供了基础。通过实现神经重用，神经形态硬件的分类准确率可提高17%，同时功耗降低40%；在多任务场景中，其训练速度加快2200%，同时泛化能力增强21%。我们的研究结果对于构建基于铁电-离子组合的可刷新硬件平台具有前景，该平台能够容纳更高效的算法和架构。
