Unsupervised feature learning in spiking neural networks using nonlinear interface dipole modulation-based synaptic devices

Recently, a three-terminal interface dipole modulation field-effect transistor (IDM FET) memory device has been proposed that leverages electric-field-induced dipole modulation at oxide/oxide interfaces. This device has been reported to exhibit a double-pulse-induced response analogous to the spike-timing-dependent plasticity (STDP) observed in biological synapses. Although the STDP behavior of the IDM FET exhibits pronounced nonlinearity, previous simulation studies have suggested that it can still be applied to unsupervised feature learning in spiking neural networks (SNNs) when combined with an additional frequency-independent (FI) depression operation. In this study, we first briefly review the nonlinear IDM response based on experimental observations and clarify that the nonlinearity is intrinsic to the IDM interface, originating from changes in the interface dipole states. We then present the synaptic weight-update model of IDM FETs employed in our SNN simulations and analyze the weight-update dynamics during feature learning using a simple single-layer SNN. Based on this analysis, we examine the optimal update conditions in terms of the balance between potentiation and depression rates. Furthermore, we evaluate feature learning on the MNIST handwritten-digit dataset using a two-layer network. Based on frequency-dependent rate-equilibrium considerations, we propose a switching FI depression/potentiation algorithm to improve feature‐learning performance, demonstrating enhanced robustness, improved classification accuracy, and reasonable tolerance to device-to-device variation.