Implementation of restricted Boltzmann machine using ferroelectric-based stochastic neurons and deterministic synapses

Restricted Boltzmann machine (RBM) is a typical stochastic neural network with wide applications in image generation and recognition. Hardware implementation of RBMs using emerging devices (e.g., memristors) is promising for high energy efficiency, but it remains challenging because the devices with distinctly different switching characteristics are required to implement the stochastic neurons and deterministic synapses of the RBM, respectively. Here, an all-ferroelectric RBM, which consists of neurons based on nanoscale ferroelectric field-effect transistors (FeFETs) exhibiting probabilistic switching behavior and synapses based on microscale ferroelectric tunnel junctions (FTJs) showing deterministic switching behavior, is demonstrated by simulation. It is shown that the all-ferroelectric RBM is capable of effectively capturing and processing essential features of input images, achieving high image reconstruction performance. Moreover, the all-ferroelectric RBM's feature extraction capability is further enhanced by optimizing the device parameters of the FeFET and the mapping coefficients in the pulse scheme. Subsequently, the optimized all-ferroelectric RBM is integrated with an FTJ-based artificial neural network for image recognition. This integrated system achieves an accuracy of ∼91.8%, which almost reaches the ∼92% benchmark of a software model, demonstrating the effectiveness of the all-ferroelectric RBM as a feature extractor. This study presents a novel hardware-based approach for developing RBMs by using ferroelectric memristors with distinct switching characteristics.