Efficient Neural Network Compression Process to Facilitate Hardware Implementation

Neural networks deliver unparalleled performance across numerous domains, yet they are often burdened by increasing parameter counts and computational costs. In scenarios where real-time or embedded device deployment is required, cloud computing may not be feasible. While various neural network compression techniques have been proposed, typically a compromise between different methods is necessary. This paper introduces a new compression method that enables simultaneous quantization and pruning. Subsequently, a low-rank decomposition is performed. This approach is applied to a Long-Short Term Memory (LSTM)-based neural network and evaluated against other compression methods using a language dataset. Our results demonstrate that incorporating the proposed method into the compression process not only is feasible but also achieves significant compression efficacy.