Stable Self-Assembled Atomic-Switch Networks for Neuromorphic Applications

Abstract

Nature inspired neuromorphic architectures are being explored as an alternative to imminent limitations of conventional complementary metal-oxide semiconductor (CMOS) architectures. Utilization of such architectures for practical applications like advanced pattern recognition tasks will require synaptic connections that are both reconfigurable and stable. Here, we report realization of stable atomic-switch networks (ASN), with inherent complex connectivity, self-assembled from percolating metal nanoparticles (NPs). The device conductance reflects the configuration of synapses which can be modulated via voltage stimulus. By controlling Relative Humidity (RH) and oxygen partial-pressure during NP deposition we obtain stochastic conductance switching that is stable over several months. Detailed characterization reveals signatures of electric-field induced atomic-wire formation within the tunnel-gaps of the oxidized percolating network. Finally we show that the synaptic structure can be reconfigured by stimulating at different repetition rates, which can be utilized as short-term to long-term memory conversion. This demonstration of stable stochastic switching in ASNs provides a promising route to hardware implementation of biological neuronal models and, as an example, we highlight possible applications in Reservoir Computing (RC). Index TermsAtomic switch networks, Clusters, Neuromorphic architecture I. INTRODUCTIONT HE astounding success of the von Neumann architecture for computers [1], as encapsulated in Moore's Law, is now meeting with fundamental limitations (physical transistor dimensions are approaching classical limits) and practical limitations (the exponential increase in research and development costs for every new process line) [2], [3]. Natural information processing systems, like the biological brain, on the other hand, can perform highly complex computational tasks like navigation, recognition and decision-making with remarkable ease and with very low energy consumption [4]. This natural computation, processing the useful data (patterns) from a multitude of sensory information, is immediate and cannot be matched by even the most-advanced supercomputers [5], [6]. Nature inspired architectures [7]-[13] are therefore currently being pursued as a disruptive alternative to the von Neumann architecture. A recent review on Neuromorphic architecture [14] and implementations can be found in Ref. [15].The alternative brain-inspired hardware approach must address three key issues simultaneously: mimic the complex biological network of neurons, replicate synaptic structures and allow implementation of standard computational algorithms [16]. Achieving all of these goals is obviously enormously challenging and will require long-term research. Nevertheless significant progress has been made towards solving several different problems, using a variety of architectures. Proposals for non-CMOS approaches include those based on networks of memristors [17]-[19], atomic switches [10], [11] and Metal Oxide Resistive Rando...

Synaptic dynamics in complex self-assembled nanoparticle networks

“…Our devices are fabricated using simple cluster deposition techniques that have been described in detail previously 28,32 . Briefly, tin clusters (∼ 8.5 nm) are deposited between Au contacts (100 µm separation) in a controlled atmosphere containing a small amount of air (P dep ∼ 1 × 10 −5 Torr) with high relative humidity, 28 resulting in the morphology shown in Fig. 1 (a).…”

“…Active switching around this conductance value can be observed for many months. 28 Fig. 2(e) summarizes the dependence of the number of events/pulse as a function of τ p and V p , and emphasises the dramatic increase in activity of the network for stronger excitation pulses (higher V p ).…”

“…This complex synaptic network could be imagined as a dynamic reservoir 12,35,36 of interconnected and interdependent synaptic 'weights'. Reservoir computing (RC) using such atomic switch networks is believed to be a powerful tool for a range of problems including pattern recognition 10,28,35,37 and various waveform regression tasks 38 .…”

“…2(a-d)]: this stability for sub-threshold voltages (V p < V T ) has been verified for extended periods (weeks) of time. 28 Every time a synapse connects/disconnects the electric field and current distribution is rearranged and subsequent synaptic activity is dependent upon the modified network state. As V p is increased beyond V T switching becomes more active and the devices reach a state in which the average conductance is ∼ 2G 0 .…”

“…It has recently been shown that these synaptic networks can be fabricated such that devices work reproducibly for several months. 28 Since the networks are complex, it is intrinsically difficult to separate out the switching of the individual elements from the observed changes of the network as a whole. Here we apply controlled voltage pulses and characterize the resultant network dynamics.…”