来自中国科学院深圳先进技术研究院的研究人员近日在Applied physics letters发表了题为“Flexible electronic synapse enabled by ferroelectric field effect transistor for robust neuromorphic computing”的文 ...
文章链接:https://doi.org/10.1063/5.0013638 基于铁电柔性晶体管的人工突触 背景 基于冯·诺依曼体系的传统观点在后摩尔时代面临着巨大的挑战,即以更低的能耗提供不断增长的计算能力。为了克服这个问题,由大脑激发的神经形态计算范式引起了极大的关注。人脑是一个高效,柔软但强大的生物计算系统,能够处理各种复杂的任务。它具有约20W极低功耗,杨氏模量低至约10 kPa,并且在各种外部干扰下都能很好地工作,这是任何人造机器无法比拟的。人脑信息传输和处理的基本单位是神经元(〜10^11)和突触(〜10^15)。作为神经元的特殊连接点,突触能够就地存储和处理信息,实现高效并行和低功耗存内计算。因此,人们付出了很多努力来开发用于模拟神经形态计算的模拟突触行为的人造电子设备。 摘要 神经形态计算具有加速高性能并行和低功耗内存中计算,人工智能和自适应学习的潜力。尽管很好地模仿了生物突触的基本功能,但是现有的人工电子突触装置仍未达到大脑的柔软性,鲁棒性和超低功耗。在这里,我们演示了一种由基于云母的铁电场效应晶体管实现的全无机柔性人工突触。该设备不仅具有出色的电脉冲调制电导率更新能力,而且还具有出色的机械柔韧性和高温可靠性,从而在应力和加热等外部干扰条件下进行强大的神经形态计算成为可能。基于其线性,可重复和稳定的长期可塑性,我们模拟了美国国家标准与技术研究院手写数字识别的人工神经网络,其准确度为94.4%。这项工作提供了一种有前途的方法,可以实现模仿大脑的灵活,低功耗,健壮和高效的神经形态计算。 Neuromorphic computing has the potential to accelerate high performance parallel and low power in-memory computation, artificial intelligence, and adaptive learning. Despite emulating the basic functions of biological synapses well, the existing artificial electronic synaptic devices have yet to match the softness, robustness, and ultralow power consumption of the brain. Here, we demonstrate an all-inorganic flexible artificial synapse enabled by a ferroelectric field effect transistor based on mica. The device not only exhibits excellent electrical pulse modulated conductance updating for synaptic functions but also shows remarkable mechanical flexibility and high temperature reliability, making robust neuromorphic computation possible under external disturbances such as stress and heating. Based on its linear, repeatable, and stable long-term plasticity, we simulate an artificial neural network for the Modified National Institute of Standards and Technology handwritten digit recognition with an accuracy of 94.4%. This work provides a promising way to enable flexible, low-power, robust, and highly efficient neuromorphic computation that mimics the brain. |