DP-GP-LVM: A Bayesian Non-Parametric Model for Learning Multivariate Dependency Structures

Authors: Andrew Lawrence, Carl Henrik Ek, Neill Campbell

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of our approach via analysis of discovered structure and superior quantitative performance on missing data imputation. We test the model on four different data sets with the aim of providing intuition to the benefit of our approach in comparison with previous models. Additional results are in appendix G.
Researcher Affiliation Academia 1Dept. of Computer Science, University of Bath, UK 2Dept. of Computer Science, University of Bristol, UK. Correspondence to: Andrew R. Lawrence <A.R.Lawrence@bath.ac.uk>.
Pseudocode No Paper describes methods textually and mathematically but does not include any pseudocode or algorithm blocks.
Open Source Code No The paper does not provide a link to the open-source code for the described methodology or explicitly state its release.
Open Datasets Yes Motion Capture Datasets Motion capture data is represented in high-dimensional vector spaces but due to the underlying structure of the motion and the human body the data resides on a much lower-dimensional manifold. These correlation structures are challenging to specify a priori making this data ideal to demonstrate our approach. Pose Track (Andriluka et al., 2018) consists of spatial image locations corresponding to an underlying human 3-D motion. In a third data set, we apply the model to a 3-D motion of a horse (Abson, 2014) shown in Fig. 5 with further details in the appendix ( G).
Dataset Splits No The paper does not specify exact training, validation, or test dataset splits (e.g., percentages, sample counts, or predefined split citations).
Hardware Specification No The paper does not specify any hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions 'Tensor Flow (Abadi et al., 2015)' but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We initialize the mean parameters µ for q(X) with the first Q principal components of the observed data Y and set all Σq = 1 2IN. The pseudo input locations Xu are initialized to a random subset of µ. The stick length parameters a and b are drawn from a standard log-normal. The allocation parameters Φ are drawn from a standard normal pushed through the soft-max function. The hyperparameters and noise precisions are initialized with draws from their log-normal priors. Finally, the shape and scale for the gamma distribution over α are initialized to their prior w1 = s1 = w2 = s2 = 1.