Empirical analysis of the divergence of Gibbs sampling based learning algorithms for restricted Boltzmann machines
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Standard
Empirical analysis of the divergence of Gibbs sampling based learning algorithms for restricted Boltzmann machines. / Fischer, Asja; Igel, Christian.
Artificial Neural Networks – ICANN 201: 20th International Conference, Thessaloniki, Greece, September 15-18, 2010, Proceedings, Part III. ed. / K. Diamantaras; W. Duch; L. S. Iliadis. Springer, 2010. p. 208-217 (Lecture notes in computer science, Vol. 6354).Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - GEN
T1 - Empirical analysis of the divergence of Gibbs sampling based learning algorithms for restricted Boltzmann machines
AU - Fischer, Asja
AU - Igel, Christian
PY - 2010
Y1 - 2010
N2 - Learning algorithms relying on Gibbs sampling based stochastic approximations of the log-likelihood gradient have become a common way to train Restricted Boltzmann Machines (RBMs). We study three of these methods, Contrastive Divergence (CD) and its refined variants Persistent CD (PCD) and Fast PCD (FPCD). As the approximations are biased, the maximum of the log-likelihood is not necessarily obtained. Recently, it has been shown that CD, PCD, and FPCD can even lead to a steady decrease of the log-likelihood during learning. Taking artificial data sets from the literature we study these divergence effects in more detail. Our results indicate that the log-likelihood seems to diverge especially if the target distribution is difficult to learn for the RBM. The decrease of the likelihood can not be detected by an increase of the reconstruction error, which has been proposed as a stopping criterion for CD learning. Weight-decay with a carefully chosen weight-decay-parameter can prevent divergence.
AB - Learning algorithms relying on Gibbs sampling based stochastic approximations of the log-likelihood gradient have become a common way to train Restricted Boltzmann Machines (RBMs). We study three of these methods, Contrastive Divergence (CD) and its refined variants Persistent CD (PCD) and Fast PCD (FPCD). As the approximations are biased, the maximum of the log-likelihood is not necessarily obtained. Recently, it has been shown that CD, PCD, and FPCD can even lead to a steady decrease of the log-likelihood during learning. Taking artificial data sets from the literature we study these divergence effects in more detail. Our results indicate that the log-likelihood seems to diverge especially if the target distribution is difficult to learn for the RBM. The decrease of the likelihood can not be detected by an increase of the reconstruction error, which has been proposed as a stopping criterion for CD learning. Weight-decay with a carefully chosen weight-decay-parameter can prevent divergence.
U2 - 10.1007/978-3-642-15825-4_26
DO - 10.1007/978-3-642-15825-4_26
M3 - Article in proceedings
SN - 978-3-642-15824-7
T3 - Lecture notes in computer science
SP - 208
EP - 217
BT - Artificial Neural Networks – ICANN 201
A2 - Diamantaras, K.
A2 - Duch, W.
A2 - Iliadis, L. S.
PB - Springer
Y2 - 15 September 2010 through 18 September 2010
ER -
ID: 33862803