Tunable Floating-Point for Artificial Neural Networks

Marta Franceschi; Alberto Nannarelli; Maurizio Valle

doi:10.1109/ICECS.2018.8617900

Tunable Floating-Point for Artificial Neural Networks

Marta Franceschi, Alberto Nannarelli, Maurizio Valle

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

323 Downloads (Pure)

Abstract

Approximate computing has emerged as a promising approach to energy-efficient design of digital systems in many domains such as digital signal processing, robotics, and machine learning. Numerous studies report that employing different data formats in Deep Neural Networks (DNNs), the dominant Machine Learning approach, could allow substantial improvements in power efficiency considering an acceptable quality for results. In this work, the application of Tunable Floating-Point (TFP) precision to DNN is presented. In TFP different precisions for different operations can be set by selecting a specific number of bits for significand and exponent in the floating-point representation. Flexibility in tuning the precision of given layers of the
neural network may result in a more power efficient computation.

Original language	English
Title of host publication	Proceedings of 25th IEEE International Conference on Electronics Circuits and Systems
Publisher	IEEE
Publication date	2018
Pages	289-292
ISBN (Print)	9781538695623
DOIs	https://doi.org/10.1109/ICECS.2018.8617900
Publication status	Published - 2018
Event	2018 IEEE 25th International Conference on Electronics, Circuits and Systems - Palais des Congrès, Bordeaux, France Duration: 9 Dec 2018 → 12 Dec 2018 Conference number: 25 https://ieeexplore.ieee.org/xpl/conhome/8599658/proceeding

Conference

Conference	2018 IEEE 25th International Conference on Electronics, Circuits and Systems
Number	25
Location	Palais des Congrès
Country/Territory	France
City	Bordeaux
Period	09/12/2018 → 12/12/2018
Internet address	https://ieeexplore.ieee.org/xpl/conhome/8599658/proceeding

Keywords

Floating-point
Power efficiency
Neural networks

Access to Document

10.1109/ICECS.2018.8617900

PID5613277Submitted manuscript, 756 KB

SFX availability

Full text

Cite this

@inproceedings{551e92f4ec9a454692cd29706f154886,

title = "Tunable Floating-Point for Artificial Neural Networks",

abstract = "Approximate computing has emerged as a promising approach to energy-efficient design of digital systems in many domains such as digital signal processing, robotics, and machine learning. Numerous studies report that employing different data formats in Deep Neural Networks (DNNs), the dominant Machine Learning approach, could allow substantial improvements in power efficiency considering an acceptable quality for results. In this work, the application of Tunable Floating-Point (TFP) precision to DNN is presented. In TFP different precisions for different operations can be set by selecting a specific number of bits for significand and exponent in the floating-point representation. Flexibility in tuning the precision of given layers of theneural network may result in a more power efficient computation.",

keywords = "Floating-point, Power efficiency, Neural networks",

author = "Marta Franceschi and Alberto Nannarelli and Maurizio Valle",

year = "2018",

doi = "10.1109/ICECS.2018.8617900",

language = "English",

isbn = "9781538695623",

pages = "289--292",

booktitle = "Proceedings of 25th IEEE International Conference on Electronics Circuits and Systems",

publisher = "IEEE",

address = "United States",

note = "2018 IEEE 25th International Conference on Electronics, Circuits and Systems, ICECS 2018 ; Conference date: 09-12-2018 Through 12-12-2018",

url = "https://ieeexplore.ieee.org/xpl/conhome/8599658/proceeding",

}

TY - GEN

T1 - Tunable Floating-Point for Artificial Neural Networks

AU - Franceschi, Marta

AU - Nannarelli, Alberto

AU - Valle, Maurizio

N1 - Conference code: 25

PY - 2018

Y1 - 2018

N2 - Approximate computing has emerged as a promising approach to energy-efficient design of digital systems in many domains such as digital signal processing, robotics, and machine learning. Numerous studies report that employing different data formats in Deep Neural Networks (DNNs), the dominant Machine Learning approach, could allow substantial improvements in power efficiency considering an acceptable quality for results. In this work, the application of Tunable Floating-Point (TFP) precision to DNN is presented. In TFP different precisions for different operations can be set by selecting a specific number of bits for significand and exponent in the floating-point representation. Flexibility in tuning the precision of given layers of theneural network may result in a more power efficient computation.

AB - Approximate computing has emerged as a promising approach to energy-efficient design of digital systems in many domains such as digital signal processing, robotics, and machine learning. Numerous studies report that employing different data formats in Deep Neural Networks (DNNs), the dominant Machine Learning approach, could allow substantial improvements in power efficiency considering an acceptable quality for results. In this work, the application of Tunable Floating-Point (TFP) precision to DNN is presented. In TFP different precisions for different operations can be set by selecting a specific number of bits for significand and exponent in the floating-point representation. Flexibility in tuning the precision of given layers of theneural network may result in a more power efficient computation.

KW - Floating-point

KW - Power efficiency

KW - Neural networks

U2 - 10.1109/ICECS.2018.8617900

DO - 10.1109/ICECS.2018.8617900

M3 - Article in proceedings

SN - 9781538695623

SP - 289

EP - 292

BT - Proceedings of 25th IEEE International Conference on Electronics Circuits and Systems

PB - IEEE

T2 - 2018 IEEE 25th International Conference on Electronics, Circuits and Systems

Y2 - 9 December 2018 through 12 December 2018

ER -

Tunable Floating-Point for Artificial Neural Networks

Abstract

Conference

Keywords

Access to Document

SFX availability

Fingerprint

Cite this