Analog memory and spike-timing-dependent plasticity characteristics of a nanoscale titanium oxide bilayer resistive switching device

Abstract

We demonstrated analog memory, synaptic plasticity, and a spike-timing-dependent plasticity (STDP) function with a nanoscale titanium oxide bilayer resistive switching device with a simple fabrication process and good yield uniformity. We confirmed the multilevel conductance and analog memory characteristics as well as the uniformity and separated states for the accuracy of conductance change. Finally, STDP and a biological triple model were analyzed to demonstrate the potential of titanium oxide bilayer resistive switching device as synapses in neuromorphic devices. By developing a simple resistive switching device that can emulate a synaptic function, the unique characteristics of synapses in the brain, e.g. combined memory and computing in one synapse and adaptation to the outside environment, were successfully demonstrated in a solid state device.

Multilevel resistive switching in TiO2/Al2O3 bilayers at low temperature

“…The specific feature of the Pt/TiO 2 /Al 2 O 3 /Pt structure is the possibility of tuning its resistance from 0.8×10 12 ·up to 6×10 5 Ω·by applying a relatively small negative voltage (U SL in Fig. 3(a, c)) to the Pt-TE.…”

“…The synapticstrength distribution provides the neural network with memory, while the synapse reconfiguration (plasticity) is responsible for learning. 2 An individual device based on a single-layer MIM structure has already been demonstrated as an element of neural networks with autonomous learning [3][4][5][6][7][8][9][10] (i.e., learning without any external control of the synaptic strength or any previous knowledge of the information to be processed). In this case, the low-resistance state of a MIM structure corresponds to a strong synaptic connection, whereas the high-resistance state corresponds to a weak synaptic connection.…”

Pattern classification by memristive crossbar circuits using ex situ and in situ training

“…The recent milestones for metal-oxide memristors, which are compatible with existing silicon technology, include demonstrations of sub-10-nm devices 16 , 10 12 -cycle endurance 17 , pico-Joule 18 and sub-ns switching 19 , and the monolithic integration of several memristive crossbar layers 20 . These milestones, in turn, revived interest in the development of memristor-based ANNs and led to numerous few-or single-memristor demonstrations of synaptic functionality and simple associative memory [21][22][23][24][25][26][27][28] , as well as the theoretical modelling of large-scale networks 10,[29][30][31][32][33] . Despite significant progress in memristor crossbar memories 20,[34][35][36][37] , memristor-based ANNs have proven to be significantly more challenging and have yet to be demonstrated.…”

Mimicking of pulse shape-dependent learning rules with a quantum dot memristor

“…[19][20][21] For example, in hippocampal neurons, potentiation (increase) of the synaptic strength is observed when the post-follows the presynaptic pulse, while depression (decrease) occurs when the pre-follows the postsynaptic pulse (asymmetric Hebbian learning). 16 This functionality can be successfully emulated with memristors 4,[22][23][24][25][26] and, empirically, it is described with exponential functions. 14,27 Depending on the synapse type (excitatory or inhibitory), potentiation and depression can also occur for a reversed order of the pre-and postsynaptic pulses (asymmetric anti-Hebbian learning).…”