TY - GEN
T1 - Adaptive importance sampling supported by a variational auto-encoder
AU - Wang, Hechuan
AU - Bugallo, Monica F.
AU - Djuric, Petar M.
N1 - Publisher Copyright:
© 2019 IEEE.
PY - 2019/12
Y1 - 2019/12
N2 - Obtaining good samples from high dimensional target distributions by adaptive importance sampling (AIS) methods to approximate target distributions is usually very hard, especially when not much is known about the target distribution. The main difficulty is in adapting the proposal distribution to become closer to the target distribution. The Variational Auto-Encoder (VAE) is a probabilistic deep computational network that can capture lower dimensional coding of high dimensional data. This property gives it a potential to capture the structure of high dimensional target distributions. In this paper, we propose an AIS method that exploits variational autoencoding to construct better proposal distributions. We demonstrate that the proposed method has better performance than the standard AIS in terms of convergence speed and goodness-of-fit when sampling from high dimensional and highly structured target distributions.
AB - Obtaining good samples from high dimensional target distributions by adaptive importance sampling (AIS) methods to approximate target distributions is usually very hard, especially when not much is known about the target distribution. The main difficulty is in adapting the proposal distribution to become closer to the target distribution. The Variational Auto-Encoder (VAE) is a probabilistic deep computational network that can capture lower dimensional coding of high dimensional data. This property gives it a potential to capture the structure of high dimensional target distributions. In this paper, we propose an AIS method that exploits variational autoencoding to construct better proposal distributions. We demonstrate that the proposed method has better performance than the standard AIS in terms of convergence speed and goodness-of-fit when sampling from high dimensional and highly structured target distributions.
KW - adaptive importance sampling
KW - population Monte Carlo
KW - variational euto-encoders
UR - https://www.scopus.com/pages/publications/85082386535
U2 - 10.1109/CAMSAP45676.2019.9022490
DO - 10.1109/CAMSAP45676.2019.9022490
M3 - Conference contribution
AN - SCOPUS:85082386535
T3 - 2019 IEEE 8th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, CAMSAP 2019 - Proceedings
SP - 619
EP - 623
BT - 2019 IEEE 8th International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, CAMSAP 2019 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 8th IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing, CAMSAP 2019
Y2 - 15 December 2019 through 18 December 2019
ER -