Encoding of Neural and Deep Learning Networks using generalized bit-planes: fostering sparsity while reducing storage demands and computational complexity

Deep neural networks (DNN) achieve high performance across many tasks. They often require extensive parameter sets and computationally intensive weight representations. The memory footprint, storage demand, and arithmetic complexity directly affect the practical deployment of DNN models, especially in resource-constrained settings. We investigate post-training quantization of DNN models based on sums of powers of two (SOPOT), resulting in model compression. This is done using the Matching Pursuits with Generalized Bit Planes (MPGBP) algorithm. The proposed approach replaces the conventional weight formats with sparse, signed power-of-two (SPTs) representations that are better suited to low-complexity implementation. We compare three approximation granularities: neuron-, layer-, and network-wise. Their trade-offs in terms of accuracy, sparsity, and entropy-based rate on VGG-16 pretrained on ImageNet-1K are evaluated. The results show that the network-wise strategy provides the best compression-performance trade-off.