A Dual-Mode ReRAM CIM Macro for Low Power Memory-Augmented Neural Networks

Learning from a few unseen samples to generalize the feature of new classes is a key challenge for real-time machine intelligence. To resolve this issue, the memory-augmented neural network (MANN) is proposed with an external memory to make use of the learned knowledge patterns to adapt to the new data [1]. A ReRAM computing-in-memory (CIM) macro based on voltage division is proposed to support the MANN operations in this paper. With the same bit-cell but different encoding schemes, the proposed macro is able to realize two different computing paradigms. The proposed complementary structure of bit-cell shows a high energy efficiency in the compute phase. And the simulation result shows that the proposed bit-cell has a good tolerance of device variations as well.