Название: Embedded Machine Learning for Cyber-Physical, IoT, and Edge Computing: Use Cases and Emerging Challenges
Автор: Sudeep Pasricha, Muhammad Shafique
Издательство: Springer
Год: 2024
Страниц: 571
Язык: английский
Формат: pdf (true)
Размер: 28.9 MB

Machine Learning (ML) has emerged as a prominent approach for achieving state-of-the-art accuracy for many data analytic applications, ranging from computer vision (e.g., classification, segmentation, and object detection in images and video), speech recognition, language translation, healthcare diagnostics, robotics, and autonomous vehicles to business and financial analysis. The driving force of the ML success is the advent of Neural Network (NN) algorithms, such as Deep Neural Networks (DNNs)/Deep Learning (DL) and Spiking Neural Networks (SNNs), with support from today’s evolving computing landscape to better exploit data and thread-level parallelism with ML accelerators.

Current trends show an immense interest in attaining the powerful abilities of NN algorithms for solving ML tasks using embedded systems with limited compute and memory resources, i.e., so-called Embedded ML. One of the main reasons is that embedded ML systems may enable a wide range of applications, especially the ones with tight memory and power/energy constraints, such as mobile systems, Internet of Things (IoT), edge computing, and cyber-physical applications. Furthermore, embedded ML systems can also improve the quality of service (e.g., personalized systems) and privacy as compared to centralized ML systems (e.g., based on cloud computing). However, state-of-the-art NN-based ML algorithms are costly in terms of memory sizes and power/energy consumption, thereby making it difficult to enable embedded ML systems.

This book consists of three volumes, and explores and identifies the most challenging issues that hinder the implementation of embedded ML systems. These issues arise from the fact that, to achieve better accuracy, the development of NN algorithms have led to state-of-the-art models with higher complexity with respect to model sizes and operations, the implications of which are discussed below:

• Massive Model Sizes: Larger NN models usually obtain higher accuracy than the smaller ones because they have a larger number of NN parameters that can learn the features from the training dataset better. However, a huge number of parameters may not be fully stored on-chip, hence requiring large-sized off-chip memory to store them and intensive off-chip memory accesses during run time. Furthermore, these intensive off-chip accesses are significantly more expensive in terms of latency and energy than on-chip operations, hence exacerbating the overall system energy.

• Complex and Intensive Operations: The complexity of operations in NN algorithms depends on the computational model and the network architecture. For instance, DNNs and SNNs have different complexity of operations since DNNs typically employ Multiply-and-Accumulate (MAC) while SNNs employ more bio-plausible operations like Leaky-Integrate-and-Fire (LIF). Besides, more complex neural architectures (e.g., residual networks) may require additional operations to accommodate the architectural variations. These complex architectures with a huge number of parameters also lead to intensive neural operations (e.g., a large number of MAC operations in DNNs), thereby requiring high computing power/energy during model execution.

In summary, achieving acceptable accuracy for the given ML applications while meeting the latency, memory, and power/energy constraints of the embedded ML systems is not a trivial task.

To address these challenges, this book discusses potential solutions from multiple design aspects, presents multiple applications that can benefit from embedded ML systems, and discusses the security, privacy, and robustness aspects of embedded ML systems. To provide a comprehensive coverage of all these different topics, which are crucial for designing and deploying embedded ML for real-world applications, this book is partitioned into three volumes. The first volume covers the Hardware Architectures, the second volume covers Software Optimizations and Hardware/Software Codesign, and the third volume presents different Use Cases and Emerging Challenges.