Artificial synapses with full synapse-like functionalities are of crucial importance for the implementation of neuromorphic computing and bioinspired intelligent systems. In particular, the development of artificial s...Artificial synapses with full synapse-like functionalities are of crucial importance for the implementation of neuromorphic computing and bioinspired intelligent systems. In particular, the development of artificial synapses with the capability to emulate multiplexed neural transmission is highly desirable, but remains challenging. In this work, we proposed a hybrid ambipolar synaptic transistor that combines two-dimensional(2D) molybdenum disulfide(Mo S_(2)) sheet and crystalline one-dimensional(1D) poly(3-hexylthiophene-2,5-diyl) polymer nanowires(P3HT NWs) as dual excitatory channels. Essential synaptic functions, including excitatory postsynaptic current, paired-pulse facilitation, synaptic potentiation and depression, and dynamic filtering were emulated using the synaptic transistor. Benefitting from the dual excitatory channels of the synaptic transistor, the device achieved a fast switch between short-term and long-term memory by altering the charge carriers in the dual channels, i.e., electrons and holes. This emulated the multiplexed neural transmission of different excitatory neurotransmitters, e.g., dopamine and noradrenaline. The plasticity-switchable artificial synapse(PSAS) simulates the task-learning process of individuals under different motivations and the impact of success or failure on task learning and memory, which promises the potential to enable complex functionalities in future neuromorphic intelligent electronics.展开更多
目的人脸超分辨率重建是特定应用领域的超分辨率问题,为了充分利用面部先验知识,提出一种基于多任务联合学习的深度人脸超分辨率重建算法。方法首先使用残差学习和对称式跨层连接网络提取低分辨率人脸的多层次特征,根据不同任务的学习...目的人脸超分辨率重建是特定应用领域的超分辨率问题,为了充分利用面部先验知识,提出一种基于多任务联合学习的深度人脸超分辨率重建算法。方法首先使用残差学习和对称式跨层连接网络提取低分辨率人脸的多层次特征,根据不同任务的学习难易程度设置损失权重和损失阈值,对网络进行多属性联合学习训练。然后使用感知损失函数衡量HR(high-resolution)图像与SR(super-resolution)图像在语义层面的差距,并论证感知损失在提高人脸语义信息重建效果方面的有效性。最后对人脸属性数据集进行增强,在此基础上进行联合多任务学习,以获得视觉感知效果更加真实的超分辨率结果。结果使用峰值信噪比(PSNR)和结构相似度(SSIM)两个客观评价标准对实验结果进行评价,并与其他主流方法进行对比。实验结果显示,在人脸属性数据集(Celeb A)上,在放大8倍时,与通用超分辨率MemNet(persistent memory network)算法和人脸超分辨率FSRNet(end-to-end learning face super-resolution network)算法相比,本文算法的PSNR分别提升约2.15 d B和1.2 d B。结论实验数据与效果图表明本文算法可以更好地利用人脸先验知识,产生在视觉感知上更加真实和清晰的人脸边缘和纹理细节。展开更多
基金supported by the National Science Fund for Distinguished Young Scholars of China (No. T2125005)the Tianjin Science Foundation for Distinguished Young Scholars (No. 19JCJQJC61000)the Shenzhen Science and Technology Project (No. JCYJ20210324121002008)。
文摘Artificial synapses with full synapse-like functionalities are of crucial importance for the implementation of neuromorphic computing and bioinspired intelligent systems. In particular, the development of artificial synapses with the capability to emulate multiplexed neural transmission is highly desirable, but remains challenging. In this work, we proposed a hybrid ambipolar synaptic transistor that combines two-dimensional(2D) molybdenum disulfide(Mo S_(2)) sheet and crystalline one-dimensional(1D) poly(3-hexylthiophene-2,5-diyl) polymer nanowires(P3HT NWs) as dual excitatory channels. Essential synaptic functions, including excitatory postsynaptic current, paired-pulse facilitation, synaptic potentiation and depression, and dynamic filtering were emulated using the synaptic transistor. Benefitting from the dual excitatory channels of the synaptic transistor, the device achieved a fast switch between short-term and long-term memory by altering the charge carriers in the dual channels, i.e., electrons and holes. This emulated the multiplexed neural transmission of different excitatory neurotransmitters, e.g., dopamine and noradrenaline. The plasticity-switchable artificial synapse(PSAS) simulates the task-learning process of individuals under different motivations and the impact of success or failure on task learning and memory, which promises the potential to enable complex functionalities in future neuromorphic intelligent electronics.
文摘目的人脸超分辨率重建是特定应用领域的超分辨率问题,为了充分利用面部先验知识,提出一种基于多任务联合学习的深度人脸超分辨率重建算法。方法首先使用残差学习和对称式跨层连接网络提取低分辨率人脸的多层次特征,根据不同任务的学习难易程度设置损失权重和损失阈值,对网络进行多属性联合学习训练。然后使用感知损失函数衡量HR(high-resolution)图像与SR(super-resolution)图像在语义层面的差距,并论证感知损失在提高人脸语义信息重建效果方面的有效性。最后对人脸属性数据集进行增强,在此基础上进行联合多任务学习,以获得视觉感知效果更加真实的超分辨率结果。结果使用峰值信噪比(PSNR)和结构相似度(SSIM)两个客观评价标准对实验结果进行评价,并与其他主流方法进行对比。实验结果显示,在人脸属性数据集(Celeb A)上,在放大8倍时,与通用超分辨率MemNet(persistent memory network)算法和人脸超分辨率FSRNet(end-to-end learning face super-resolution network)算法相比,本文算法的PSNR分别提升约2.15 d B和1.2 d B。结论实验数据与效果图表明本文算法可以更好地利用人脸先验知识,产生在视觉感知上更加真实和清晰的人脸边缘和纹理细节。
