IC-279GP-LVM produces less structured latent representations and lower generative-classification accuracy than QEP-LVM on oil flow and MNIST

Chukwudi Paul Obite, Zhi Chang, Keyan Wu, Shiwei Lan

SourceBayesian Regularization of Latent Representation

The paper applies the standard Bayesian GP-LVM (q=2) to the oil flow and MNIST datasets and reports its generative-classification performance and latent-space geometry. On oil flow, GP-LVM achieves 0.68 ± 0.05 accuracy and 0.83 ± 0.03 AUC over 10 random seeds. On MNIST, it achieves 0.965 ± 0.012 accuracy and 0.981 ± 0.0062 AUC. The 2-D latent projections show GP-LVM yielding less concentrated digit clusters (e.g. segregated but overlapping 6/7 groups) compared to the tighter, more separated clusters from QEP-LVM at q=1.5.

Evidence

correlational

Key metric

oil flow: acc 0.68 ± 0.05, auc 0.83 ± 0.03, ari 0.34 ± 0.07, nmi 0.37 ± 0.07; MNIST: acc 0.965 ± 0.012, auc 0.981 ± 0.0062, ari 0.923 ± 0.025, nmi 0.938 ± 0.019

Caveat

Results are from the authors' own training runs with 25 (oil flow) or 128 (MNIST) inducing points and a specific ARD SE kernel; the paper does not release the trained GP-LVM checkpoints.

Model

GP-LVM

Datasets

MNIST [eval], Swiss Roll [eval]

Methods

t-SNE [eval]

Extraction

automatic-extraction