Score Matching Training for Deep Energy-Based Latent Variable Model 


Vol. 48,  No. 12, pp. 1549-1558, Dec.  2023
10.7840/kics.2023.48.12.1549


PDF
  Abstract

In this paper, we present a learning technique using score matching (SM) for the recently proposed deep energy-based latent variable model (DELVM). The energy function of DELVM has continuous inputs and hidden layers and are defined by deep neural networks. Based on this energy function, the objective function of SM-based DELVM learning is derived through Fisher divergence and presents a parameter update method using the gradient descent algorithm. In an image recognition experiment for unsupervised feature learning using Fashion MNIST and CIFAR10 data, the proposed technique demonstrates an improved performance than contrastive divergence-based DELVM and existing models. In addition, abnormal detection experiments using ECG and CIFAR10 data also show more effective F1-score over existing models.

  Statistics
Cumulative Counts from November, 2022
Multiple requests among the same browser session are counted as one view. If you mouse over a chart, the values of data points will be shown.


  Related Articles
  Cite this article

[IEEE Style]

G. Peng and D. K. Kim, "Score Matching Training for Deep Energy-Based Latent Variable Model," The Journal of Korean Institute of Communications and Information Sciences, vol. 48, no. 12, pp. 1549-1558, 2023. DOI: 10.7840/kics.2023.48.12.1549.

[ACM Style]

Guo Peng and Doon Kook Kim. 2023. Score Matching Training for Deep Energy-Based Latent Variable Model. The Journal of Korean Institute of Communications and Information Sciences, 48, 12, (2023), 1549-1558. DOI: 10.7840/kics.2023.48.12.1549.

[KICS Style]

Guo Peng and Doon Kook Kim, "Score Matching Training for Deep Energy-Based Latent Variable Model," The Journal of Korean Institute of Communications and Information Sciences, vol. 48, no. 12, pp. 1549-1558, 12. 2023. (https://doi.org/10.7840/kics.2023.48.12.1549)
Vol. 48, No. 12 Index

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)






pISSN : 1226 - 4717
eISSN : 2287 - 3880