A cost-based optimizer for gradient descent optimization

Research output: Chapter in Book/Report/Conference proceedingConference contribution

13 Citations (Scopus)

Abstract

As the use of machine learning (ML) permeates into diverse application domains, there is an urgent need to support a declarative framework for ML. Ideally, a user will specify an ML task in a high-level and easy-to-use language and the framework will invoke the appropriate algorithms and system configurations to execute it. An important observation towards designing such a framework is that many ML tasks can be expressed as mathematical optimization problems, which take a specific form. Furthermore, these optimization problems can be efficiently solved using variations of the gradient descent (GD) algorithm. Thus, to decouple a user specification of an ML task from its execution, a key component is a GD optimizer. We propose a cost-based GD optimizer that selects the best GD plan for a given ML task. To build our optimizer, we introduce a set of abstract operators for expressing GD algorithms and propose a novel approach to estimate the number of iterations a GD algorithm requires to converge. Extensive experiments on real and synthetic datasets show that our optimizer not only chooses the best GD plan but also allows for optimizations that achieve orders of magnitude performance speed-up.

Original languageEnglish
Title of host publicationSIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data
PublisherAssociation for Computing Machinery
Pages977-992
Number of pages16
VolumePart F127746
ISBN (Electronic)9781450341974
DOIs
Publication statusPublished - 9 May 2017
Event2017 ACM SIGMOD International Conference on Management of Data, SIGMOD 2017 - Chicago, United States
Duration: 14 May 201719 May 2017

Other

Other2017 ACM SIGMOD International Conference on Management of Data, SIGMOD 2017
CountryUnited States
CityChicago
Period14/5/1719/5/17

Fingerprint

Learning systems
Costs
Mathematical operators
Specifications
Experiments

ASJC Scopus subject areas

  • Software
  • Information Systems

Cite this

Kaoudi, Z., Quiané-Ruiz, J. A., Thirumuruganathan, S., Chawla, S., & Agrawal, D. (2017). A cost-based optimizer for gradient descent optimization. In SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data (Vol. Part F127746, pp. 977-992). Association for Computing Machinery. https://doi.org/10.1145/3035918.3064042

A cost-based optimizer for gradient descent optimization. / Kaoudi, Zoi; Quiané-Ruiz, Jorge Arnulfo; Thirumuruganathan, Saravanan; Chawla, Sanjay; Agrawal, Divy.

SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data. Vol. Part F127746 Association for Computing Machinery, 2017. p. 977-992.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Kaoudi, Z, Quiané-Ruiz, JA, Thirumuruganathan, S, Chawla, S & Agrawal, D 2017, A cost-based optimizer for gradient descent optimization. in SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data. vol. Part F127746, Association for Computing Machinery, pp. 977-992, 2017 ACM SIGMOD International Conference on Management of Data, SIGMOD 2017, Chicago, United States, 14/5/17. https://doi.org/10.1145/3035918.3064042
Kaoudi Z, Quiané-Ruiz JA, Thirumuruganathan S, Chawla S, Agrawal D. A cost-based optimizer for gradient descent optimization. In SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data. Vol. Part F127746. Association for Computing Machinery. 2017. p. 977-992 https://doi.org/10.1145/3035918.3064042
Kaoudi, Zoi ; Quiané-Ruiz, Jorge Arnulfo ; Thirumuruganathan, Saravanan ; Chawla, Sanjay ; Agrawal, Divy. / A cost-based optimizer for gradient descent optimization. SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data. Vol. Part F127746 Association for Computing Machinery, 2017. pp. 977-992
@inproceedings{f16b084fb7fd4aa9bcace5d772049dc3,
title = "A cost-based optimizer for gradient descent optimization",
abstract = "As the use of machine learning (ML) permeates into diverse application domains, there is an urgent need to support a declarative framework for ML. Ideally, a user will specify an ML task in a high-level and easy-to-use language and the framework will invoke the appropriate algorithms and system configurations to execute it. An important observation towards designing such a framework is that many ML tasks can be expressed as mathematical optimization problems, which take a specific form. Furthermore, these optimization problems can be efficiently solved using variations of the gradient descent (GD) algorithm. Thus, to decouple a user specification of an ML task from its execution, a key component is a GD optimizer. We propose a cost-based GD optimizer that selects the best GD plan for a given ML task. To build our optimizer, we introduce a set of abstract operators for expressing GD algorithms and propose a novel approach to estimate the number of iterations a GD algorithm requires to converge. Extensive experiments on real and synthetic datasets show that our optimizer not only chooses the best GD plan but also allows for optimizations that achieve orders of magnitude performance speed-up.",
author = "Zoi Kaoudi and Quian{\'e}-Ruiz, {Jorge Arnulfo} and Saravanan Thirumuruganathan and Sanjay Chawla and Divy Agrawal",
year = "2017",
month = "5",
day = "9",
doi = "10.1145/3035918.3064042",
language = "English",
volume = "Part F127746",
pages = "977--992",
booktitle = "SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data",
publisher = "Association for Computing Machinery",

}

TY - GEN

T1 - A cost-based optimizer for gradient descent optimization

AU - Kaoudi, Zoi

AU - Quiané-Ruiz, Jorge Arnulfo

AU - Thirumuruganathan, Saravanan

AU - Chawla, Sanjay

AU - Agrawal, Divy

PY - 2017/5/9

Y1 - 2017/5/9

N2 - As the use of machine learning (ML) permeates into diverse application domains, there is an urgent need to support a declarative framework for ML. Ideally, a user will specify an ML task in a high-level and easy-to-use language and the framework will invoke the appropriate algorithms and system configurations to execute it. An important observation towards designing such a framework is that many ML tasks can be expressed as mathematical optimization problems, which take a specific form. Furthermore, these optimization problems can be efficiently solved using variations of the gradient descent (GD) algorithm. Thus, to decouple a user specification of an ML task from its execution, a key component is a GD optimizer. We propose a cost-based GD optimizer that selects the best GD plan for a given ML task. To build our optimizer, we introduce a set of abstract operators for expressing GD algorithms and propose a novel approach to estimate the number of iterations a GD algorithm requires to converge. Extensive experiments on real and synthetic datasets show that our optimizer not only chooses the best GD plan but also allows for optimizations that achieve orders of magnitude performance speed-up.

AB - As the use of machine learning (ML) permeates into diverse application domains, there is an urgent need to support a declarative framework for ML. Ideally, a user will specify an ML task in a high-level and easy-to-use language and the framework will invoke the appropriate algorithms and system configurations to execute it. An important observation towards designing such a framework is that many ML tasks can be expressed as mathematical optimization problems, which take a specific form. Furthermore, these optimization problems can be efficiently solved using variations of the gradient descent (GD) algorithm. Thus, to decouple a user specification of an ML task from its execution, a key component is a GD optimizer. We propose a cost-based GD optimizer that selects the best GD plan for a given ML task. To build our optimizer, we introduce a set of abstract operators for expressing GD algorithms and propose a novel approach to estimate the number of iterations a GD algorithm requires to converge. Extensive experiments on real and synthetic datasets show that our optimizer not only chooses the best GD plan but also allows for optimizations that achieve orders of magnitude performance speed-up.

UR - http://www.scopus.com/inward/record.url?scp=85021237387&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85021237387&partnerID=8YFLogxK

U2 - 10.1145/3035918.3064042

DO - 10.1145/3035918.3064042

M3 - Conference contribution

VL - Part F127746

SP - 977

EP - 992

BT - SIGMOD 2017 - Proceedings of the 2017 ACM International Conference on Management of Data

PB - Association for Computing Machinery

ER -