Because of the continual advancements in the artificial
intelligence
technology, its practical applications (e.g., in computer vision,
health care, and pattern recognition) are expanding. Various architectures
integrating memory cells and transistors have been used to demonstrate
artificial synaptic arrays. However, designing memory cells with superior
analogue switching is a major challenge in the development of neuromorphic
systems. We evaluated a TiN/AlN/Cu/AlN/Pt memristive synapse for neuromorphic
computing. The synaptic device exhibited multilevel characteristics
with varying reset stop voltages ranging from −2.7 to −2.2
V. The device also exhibited highly stable reparative 200 potentiation
and depression cycles with high nonlinearity values of approximately
2.39 for potentiation and −2.19 for depression. The device
further exhibited highly stable DC endurance over 1000 cycles, AC
pulse endurance (1 M), and stable retention (104 s) at
100 °C without any degradation. The experimental potentiation
and depression data were used to train a Hopfield neural network to
effectively identify input images with a resolution of 28 × 28
pixels, representing 784 synapses used in the simulation process.
The training resulted in an accuracy of >91% in 28 epochs. Therefore,
the newly designed device exhibits promising features, such as high
linearity and accuracy, and is thus highly suitable for neuromorphic
computing purposes.