Home // PESARO 2017, The Seventh International Conference on Performance, Safety and Robustness in Complex Systems and Applications // View article

A Survey on CNN and RNN Implementations

Authors:
Javier Hoffmann
Osvaldo Navarro Guzmán
Florian Kästner
Benedikt Janßen
Michael Hübner

Keywords: deep learning, convolutional, recurrent, neural network

Abstract:
Deep Neural Networks (DNNs) are widely used for complex applications, such as image and voice processing. Two varieties of DNNs, namely Convolutional Neuronal Networks (CNNs) and Recurrent Neuronal Networks (RNNs), are particularly popular regarding recent success for industrial applications. While CNNs are typically used for computer vision applications like object recognition, RNNs are well suited for time variant problems due to their recursive structure. Even though CNNs and RNNs belong to the family of DNNs, their implementation shows substantial differences. Besides more common Central Processing Unit (CPU) and Graphic processing Unit (GPU) implementations, Field Programmable Gate Array (FPGA) implementations offer great potential. Recent evaluations have shown significant benefits of FPGA implementations of DNNs over CPUs and GPUs. In this paper, we compare current FPGA implementations of CNNs and RNNs and analyze their optimizations. With this, we provide insights regarding the specific benefits and drawbacks of recent FPGA implementations of DNNs.

Pages: 33 to 39

Copyright: Copyright (c) IARIA, 2017

Publication date: April 23, 2017

Published in: conference

ISSN: 2308-3700

ISBN: 978-1-61208-549-4

Location: Venice, Italy

Dates: from April 23, 2017 to April 27, 2017