Hybrid Parallel Inference for Hierarchical Dirichlet Processes

Tsukasa OMOTO; Koji EGUCHI; Shotaro TORA

doi:10.1587/transinf.E97.D.815

Open Access
Hybrid Parallel Inference for Hierarchical Dirichlet Processes

Tsukasa OMOTO, Koji EGUCHI, Shotaro TORA

Full Text Views

110

Share
Cite this

Free PDF (1.4MB)

Summary :

The hierarchical Dirichlet process (HDP) can provide a nonparametric prior for a mixture model with grouped data, where mixture components are shared across groups. However, the computational cost is generally very high in terms of both time and space complexity. Therefore, developing a method for fast inference of HDP remains a challenge. In this paper, we assume a symmetric multiprocessing (SMP) cluster, which has been widely used in recent years. To speed up the inference on an SMP cluster, we explore hybrid two-level parallelization of the Chinese restaurant franchise sampling scheme for HDP, especially focusing on the application to topic modeling. The methods we developed, Hybrid-AD-HDP and Hybrid-Diff-AD-HDP, make better use of SMP clusters, resulting in faster HDP inference. While the conventional parallel algorithms with a full message-passing interface does not benefit from using SMP clusters due to higher communication costs, the proposed hybrid parallel algorithms have lower communication costs and make better use of the computational resources.

Publication: IEICE TRANSACTIONS on Information Vol.E97-D No.4 pp.815-820

Publication Date: 2014/04/01

Publicized

Online ISSN: 1745-1361

DOI: 10.1587/transinf.E97.D.815

Type of Manuscript: Special Section LETTER (Special Section on Data Engineering and Information Management)

Category

Authors

Tsukasa OMOTO
  Kobe University
Koji EGUCHI
  Kobe University
Shotaro TORA
  NTT Corporation

Keyword

hierarchical dirichlet process, topic models, parallelization

Cite this

Copy

Tsukasa OMOTO, Koji EGUCHI, Shotaro TORA, "Hybrid Parallel Inference for Hierarchical Dirichlet Processes" in IEICE TRANSACTIONS on Information, vol. E97-D, no. 4, pp. 815-820, April 2014, doi: 10.1587/transinf.E97.D.815.
Abstract: The hierarchical Dirichlet process (HDP) can provide a nonparametric prior for a mixture model with grouped data, where mixture components are shared across groups. However, the computational cost is generally very high in terms of both time and space complexity. Therefore, developing a method for fast inference of HDP remains a challenge. In this paper, we assume a symmetric multiprocessing (SMP) cluster, which has been widely used in recent years. To speed up the inference on an SMP cluster, we explore hybrid two-level parallelization of the Chinese restaurant franchise sampling scheme for HDP, especially focusing on the application to topic modeling. The methods we developed, Hybrid-AD-HDP and Hybrid-Diff-AD-HDP, make better use of SMP clusters, resulting in faster HDP inference. While the conventional parallel algorithms with a full message-passing interface does not benefit from using SMP clusters due to higher communication costs, the proposed hybrid parallel algorithms have lower communication costs and make better use of the computational resources.
URL: https://globals.ieice.org/en_transactions/information/10.1587/transinf.E97.D.815/_p

Copy

@ARTICLE{e97-d_4_815,
author={Tsukasa OMOTO, Koji EGUCHI, Shotaro TORA, },
journal={IEICE TRANSACTIONS on Information},
title={Hybrid Parallel Inference for Hierarchical Dirichlet Processes},
year={2014},
volume={E97-D},
number={4},
pages={815-820},
abstract={The hierarchical Dirichlet process (HDP) can provide a nonparametric prior for a mixture model with grouped data, where mixture components are shared across groups. However, the computational cost is generally very high in terms of both time and space complexity. Therefore, developing a method for fast inference of HDP remains a challenge. In this paper, we assume a symmetric multiprocessing (SMP) cluster, which has been widely used in recent years. To speed up the inference on an SMP cluster, we explore hybrid two-level parallelization of the Chinese restaurant franchise sampling scheme for HDP, especially focusing on the application to topic modeling. The methods we developed, Hybrid-AD-HDP and Hybrid-Diff-AD-HDP, make better use of SMP clusters, resulting in faster HDP inference. While the conventional parallel algorithms with a full message-passing interface does not benefit from using SMP clusters due to higher communication costs, the proposed hybrid parallel algorithms have lower communication costs and make better use of the computational resources.},
keywords={},
doi={10.1587/transinf.E97.D.815},
ISSN={1745-1361},
month={April},}

Copy

TY - JOUR
TI - Hybrid Parallel Inference for Hierarchical Dirichlet Processes
T2 - IEICE TRANSACTIONS on Information
SP - 815
EP - 820
AU - Tsukasa OMOTO
AU - Koji EGUCHI
AU - Shotaro TORA
PY - 2014
DO - 10.1587/transinf.E97.D.815
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E97-D
IS - 4
JA - IEICE TRANSACTIONS on Information
Y1 - April 2014
AB - The hierarchical Dirichlet process (HDP) can provide a nonparametric prior for a mixture model with grouped data, where mixture components are shared across groups. However, the computational cost is generally very high in terms of both time and space complexity. Therefore, developing a method for fast inference of HDP remains a challenge. In this paper, we assume a symmetric multiprocessing (SMP) cluster, which has been widely used in recent years. To speed up the inference on an SMP cluster, we explore hybrid two-level parallelization of the Chinese restaurant franchise sampling scheme for HDP, especially focusing on the application to topic modeling. The methods we developed, Hybrid-AD-HDP and Hybrid-Diff-AD-HDP, make better use of SMP clusters, resulting in faster HDP inference. While the conventional parallel algorithms with a full message-passing interface does not benefit from using SMP clusters due to higher communication costs, the proposed hybrid parallel algorithms have lower communication costs and make better use of the computational resources.
ER -