Resistive Memory Technologies as Dynamic Elements for Neuromorphic Computing

Hardware spiking neural networks (SNNs) and emerging unconventional computing concepts hold the great promise to build future low-power computing systems. The co-development of novel efficient neuromorphic hardware and brain-inspired learning protocols aims to deliver on this promise: it will bring a paradigm shift towards edge computing, where information processing is performed close to the data generated by portable and IoT devices, and will enable applications relying on smart analysis of sensory signals in real time.<br/><br/>Recently, the research in the field of neuromorphic computing has been receiving a significant boost by exploiting the physics of the broad class of memristive devices as new hardware building blocks enabling neural behavior, and then substituting or complementing CMOS circuits. These devices can support ‘<i>in memory computing</i>’ in neural networks, thus avoiding the energy cost associated with data transfer between the memory and the processor, and then surpassing the limit of the von-Neumann computing architecture. Furthermore, as our brain uses neuronal and synaptic dynamics as working memory, it is desired that also neuromorphic chips compute with dynamics, encompassing devices and circuits with useful transients or evolution in time.<br/><br/>Our work focuses on resistive switching memory (RRAM), i.e. metal/insulator/metal devices that undergo reversible resistance change upon voltage application. RRAM devices, also depending on the used material stacks and programming strategies, allow the engineering of many functionalities for neuromorphic computing, such as analog dynamics response to input stimuli, and controllable stability of the device resistance states over various time scales. Therefore, RRAMs can act as volatile or non-volatile dynamic memory elements mimicking the short/long term plasticity of synapses in nervous system, or even as stochastic and non-linear elements in neuronal units. In our work, we engineered both analogue non-volatile HfO_x-RRAM as synaptic nodes in SNNs and Ag/SiO_x devices with tunable retention properties (from µs to seconds range), and exhibiting a conductance modulation as a function of input stimuli.<br/><br/>In an alternative approach, we developed analogue nonlinear dynamical circuits, exploiting the programmable nonlinearity of non-volatile HfO₂ RRAM devices. In particular, HfO₂-RRAMs can be programmed in different resistance states exhibiting a different nonlinearity of the current-voltage characteristic, which is used to generate complex circuit oscillations. Moreover, we have demonstrated that the circuit can be tuned from periodic (ordered) and chaotic behavior through the modulation of an input signal. In this manner, it can be used as a processing machine expanding the dimensionality of the input and enabling subsequent linear classification, in a reservoir computing scheme.<br/>To summarize, in this talk we will show how various types of RRAMs, relying on different material stacks and mechanisms, can be used for computing by exploiting their intrinsic dynamics.<br/>[<i>This work is partially supported by the Horizon 2020 EU MeM-Scales project, under the grant n. 871371]. </i>