DANNA 2: Dynamic Adaptive Neural Network Arrays

Following from the original Dynamic Adaptive Neural Network Array (DANNA) model, we propose a new digital neuromorphic architecture named DANNA 2. Through this paper, we introduce our new hardware design and software simulator, and we explain how DANNA2 can improve network density, training convergence, and element performance as compared to the DANNA model. We propose two network arrays types catering to FPGA and VLSI use cases. Using simulation results for both control and classification problems, we show improved training convergence and simulation performance.