-A A +A
The progress of deep learning in recent years—relying on deep neural networks—has driven an incredible progress in artificial intelligence [1]. Computers can now outperform professional go and poker players, recognize images better than humans in many situations, and convincingly tackle complex tasks such as text translations. These advances are changing many professional fields and might even transform our societies. Nevertheless deep learning comes with challenges. One of them is its important energy consumption [2]. In the widely publicized first victory of a computer against a professional Go player, the computer cluster consumed hundreds and thousands of watts [3], whereas a human player played only using his brain, which consumed only 20 W. Despite considerable progress since this game [4], the energy cost of operating deep neural networks means that they are typically run in data centers and not on consumer-embedded devices. This nonlocality adds to their power consumption the cost of transferring data between data center and users and raises major privacy concerns. The energy consumption of data centers is also starting to become a serious environmental concern.These considerations drive a considerable effort to develop specialized hardware for implementing deep learning. This chapter serves as an introduction to the next two chapters, which show how memristive devices can be a key asset in this quest. In the first section we introduce the general concepts of deep neural networks and their modern evolution. This section does not intend to be a comprehensive textbook about this topic, as can be found …
Woodhead Publishing
Publication date: 
12 Jun 2020

Damien Bipin Rajendran Querlioz, Sabina Spiga, Abu Sebastian

Biblio References: 
Pages: 313
Memristive Devices for Brain-Inspired Computing: From Materials, Devices, and Circuits to Applications-Computational Memory, Deep Learning, and Spiking Neural Networks