A new hardware implementation approach of BNNs based on nonlinear 2T2R synaptic cell

Abstract

For the first time, we propose a new hardware implementation approach which can utilize the non-linear synaptic cells to build a Binarized-Neural-Networks (BNNs) for online training. A 2T2R-based synaptic cell is designed and demonstrated by the fabricated RRAM array to achieve the basic functions of synapse in BNNs: binary weight (sign ( W)) reading and analog weight updating (W+ΔW). The performance of BNNs based on 2T2R synaptic cells is evaluated by MNIST, and the recognition accuracy of 97.4% can be achieved. A novel refresh operation is proposed to enhance the network performance.