Quantized neural network design under weight capacity constraint
This work addresses hardware efficiency for deep learning practitioners by providing a guideline (effective compression ratio) to balance network complexity and weight precision, though it is incremental as it builds on existing quantization and scaling methods.
The study investigated the trade-off between network size scaling and weight quantization for hardware optimization, finding that weight quantization is generally more effective than reducing network size for compressing neural networks under hardware constraints.
The complexity of deep neural network algorithms for hardware implementation can be lowered either by scaling the number of units or reducing the word-length of weights. Both approaches, however, can accompany the performance degradation although many types of research are conducted to relieve this problem. Thus, it is an important question which one, between the network size scaling and the weight quantization, is more effective for hardware optimization. For this study, the performances of fully-connected deep neural networks (FCDNNs) and convolutional neural networks (CNNs) are evaluated while changing the network complexity and the word-length of weights. Based on these experiments, we present the effective compression ratio (ECR) to guide the trade-off between the network size and the precision of weights when the hardware resource is limited.