دانلود کتاب شبکه‌های عصبی با فشرده‌سازی مدل

With the swift development of information technology, cloud computing with centralized data processing cannot meet the needs of applications that require processing massive amounts of data, and they can only be effectively used when privacy requires the data to remain at the front-end device. Thus, edge computing has become necessary to handle the data from embedded devices. Intelligent edge devices benefit many requirements within real-time unmanned aerial systems, industrial systems, and privacy-preserving applications. In recent years, deep learning has been applied to different applications, dramatically improving many artificial intelligence (AI) tasks. However, the incomparable accuracy of deep learning models is achieved by paying the cost of hungry memory consumption and high computational complexity, which significantly impedes their deployment in edge devices with low memory resources. For example, the VGG- 16 network can achieve 92.7% top-5 test accuracy on image classification tasks with the ImageNet dataset. Still, the entire network contains about 140 million 32- bit floating-point parameters, requiring more than 500 megabytes of storage space and performing .1.6 × 1010 floating-point operations. Yet, FPGA-based embedded devices typically have only a few thousand compute units, which cannot handle the millions of floating-point operations in standard deep neural network models. On the other hand, complex neural networks are often accompanied by slower computing speed and longer inference time, which are not allowed in applications with strict latency requirements, such as vehicle detection and tracking. Therefore, a natural thought is to perform model compression and acceleration in neural networks without significantly decreasing the model performance. This book introduces the significant advancements of neural networks with model compression. While quantized operations can enhance the efficiency of neural networks, they typically result in a decrease in performance. In the last 5 years, many methods have been introduced to improve the performance of quantized neural networks. To better review these methods, we focus on six aspects: gradient approximation, quantization, structural design, loss design, optimization, and neural architecture search. We also review the applications of neural networks with model compression in visual and audio analysis. There are also other model compression techniques, such as model compression with network pruning, widely used in edge computing, which we introduce for completeness in this book. From our previous studies, network pruning and quantized neural networks can be used simultaneously to complement each other, whereas network pruning on quantized neural networks can further compress models and improve the generalization ability for many downstream applications.