Auto-associative memories based on recurrent multilayer perceptrons and sparsely interconnected neural networks

摘要

Both quantized neural networks (QNNs) and sparsely interconnected neural networks (SINNs) are suitable for hardware implementation. However, quantized parameters and sparsely interconnected structure decrease the capabilities of QNNs and SINNs, respectively. In this report, we propose associative memories composed of recurrent multilayer perceptrons (RMLPs) with 3-valued weights and SINNs to improve their capabilities at a low cost.