Bienenstock, Cooper, and Munro Learning Rules Realized in Second‐Order Memristors with Tunable Forgetting Rate

Abstract

Memristors with synaptic functions are very promising for developing artificial neural networks. Compared with the extensively reported spike-timingdependent plasticity (STDP), Bienenstock, Cooper, and Munro (BCM) learning rules, the most accurate model of the synaptic plasticity to date, are more compatible with the neural computing system; however, the progress in the realization of the BCM rules has been quite limited. The realized BCM rules so far mostly performs just the spike-rate-dependent plasticity (SRDP), however, without a tunable sliding frequency threshold, because the memristors used to realize the BCM rules do not have tunable forgetting rates. In this work, the BCM rules with a tunable sliding frequency threshold are biorealistically implemented in SrTiO 3 -based second-order memristors; the forgetting rate of the memristors is tuned by engineering the electrode/oxide interface through controlling the electrode composition. The approach of this work is precise and efficient, and the biorealistic implementation of the BCM rules in memristors improves the efficiency of the neural network for the artificial intelligent system.

Toward a generalized Bienenstock-Cooper-Munro rule for spatiotemporal learning via triplet-STDP in memristive devices

“…In the BCM framework, the high/low spike rate of a train can result in the potentiation/depression of the synaptic weight depending on whether the spike rate is higher than a threshold (θ) [30][31][32][33] . For the memristor-based artificial synapse, several groups have demonstrated BCM rules using rate-based presynaptic spikes, which have led to advances in the field [34][35][36] . These results show that the absolute change in the synaptic weight (i.e., the conductance change of the memristor, |ΔG c |) has a monotonic dependence on the spike rates in both the depression region (ΔG c < 0) and potentiation region (ΔG c > 0).…”

“…However, such a monotonic change is different from the original BCM rule in neurobiology; that is, there should exist non-monotonic behavior (i.e., an enhanced depression effect (EDE)) in the depression region [30][31][32][33]37,38 . Additionally, previous memristor studies lack the following essential features: first, the lack of a multiplicative term between presynaptic and postsynaptic activities, and second the short-term modification [34][35][36] . This also marks a significant inconsistency with the biological BCM learning.…”

Memristive Synapses and Neurons for Bioinspired Computing

“…www.advelectronicmat.de be utilized to mimic LTP with forgetting effect, [159,[161][162][163] STP to LTP transformation, [157,160,[164][165][166][167][168][169][170] metaplasticitry, [112,[171][172][173][174][175] and triplet-STDP. [176][177][178] The device structure, size, switching performance, and emulated synaptic plasticity of the representative memristive devices are summarized in Table 1.…”

“…A similar approach has been adopted to biorealistically implement STDP learning rules in the biomolecular memristive devices. [157] Very recently, Xiong et al [163] realized the tunable forgetting rate by engineering the electrode/oxide interface by controlling the electrode composition in STO memristive devices, as reported in Figure 10d,e. Lu and his coworkers demonstrated BCM rules in Pd/WO x /W second-order memristive devices.…”

“…Therefore, the transformation from STP to LTP is realized by controlling the stimuli frequency. [163] Copyright 2019, Wiley-VCH. Figure 11b reports the STP to LTP transformation observed in Pt/WO x /Ti device by controlling the input numbers by Lin et al [161] In addition to two-terminal devices, three-terminal memristive devices have also been reported to realize conversion from STP to LTP.…”