Contents I. Introduction
Neuromorphic computing, which mimics the working of human brain, has seen a renewed attention since the advent of nano-technologies and the raise in demand for the automation [1], [2]. It is considered as the demand hardware for future artificial intelligence, autonomous driving systems, and complex analysis based tasks, such as image processing, signal processing, data-sensing and cognitive computing [3]. Both, hardware and software based approaches for implementing neuromorphic computing platforms are being progressively carried out [4], [5]. Using memristive devices, hardware based neuromorphic systems gained momentum in the last couple of years. RRAM based technologies are of particular interest due to their significant benefits ascribed to their CMOS compatibility [6]. To integrate RRAM devices in neuromorphic circuits, new methodologies improving the analog performance of filamentary RRAM arrays are required [7]. In this work, we utilize the HfO2 based 1T-1R RRAM devices to exhibit the analog performance of synapses. Instead of implementing discrete resistance states of binary devices into learning algorithms, the switching probability of the devices is employed to the synaptical activation function and provide a digital to analog conversion for synaptic information processing. In this respect, the device-to-device and pulse number dependent variability of the memristive devices is evaluated. Additionally, the performance and reliability of the devices used for neuromorphic computing are demonstrated through conventional endurance and retention measurements.
