Riassunto analitico
The conventional von Neumann computer architectures are now facing severe limits, among which the low energy efficiency especially when facing with non-structured or noisy data. Conversely, Biological Neural Networks (BNNs) show unmatched performance when dealing with these types of data in terms of both processing capabilities and energy consumption. Early works on artificial neural networks led to the perceptron and nowadays to Deep Neural Networks (DNNs), whose processing capabilities are at the pinnacle of what today is possible by using conventional architectures. However, strongly reducing their power consumption is one of the main challenges, that could be tackled only by unconventional circuit architectures. From this perspective, Spiking Neural Networks (SNNs) tend to better capture the typical features of biological networks, potentially offering significant improvements. In such architectures, bio-inspired unsupervised learning mechanisms such as the Spike-Timing dependent plasticity (STDP) play a central role. Recently, many new devices have been considered to enable hardware implementation of SNNs. Among the most promising, the Resistive Random-Access Memory (RRAM) offer significant benefits and a good degree of maturity. In this work, commercial-grade packaged RRAM devices are characterized in both quasi-static and pulsed regimes to test for their synaptic behavior. The results of the experimental characterization are exploited to calibrate a RRAM compact model (UniMORE model), which enables the circuit simulation of advanced neurosynaptic circuits. Specifically, a novel neural architecture is devised and simulated that can support the simultaneous presence of synapses that adopt different learning rules (such as STDP and non-STDP), thus mimicking in-silico one of the main features found in biological systems. Results show that such commercial-grade RRAM devices are suitable to be exploited as synapses in such innovative architectures, and demonstrate that the technology allows realizing neurosynaptic circuits that support many learning rules simultaneously.
|
