BATMANN: A Binarized-All-Through Memory-Augmented Neural Network for Efficient In-Memory Computing

Abstract

The traditional von Neumann architecture suffers from heavy data traffic between processing and memory units, which incurs high power and latency. To cope with the booming use of neural networks on edge devices, a promising way is to perform in-memory computing through exploiting the next-generation memristive devices. This work proposes a 2-level resistive random-access memory (RRAM)-based memory-augmented neural network (MANN), named binarized-all-through MANN (BATMANN), that is end-to-end trainable and allows both the controller and memory to be seamlessly integrated onto RRAM crossbars. Experiments then show the superiority of BATMANN in doing few-shot learning with high accuracy and robustness.