oi-VAE: Output Interpretable VAEs for Nonlinear Group Factor Analysis

Authors: Samuel K. Ainsworth, Nicholas J. Foti, Adrian K. C. Lee, Emily B. Fox

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the oi-VAE on motion capture and magnetoencephalography datasets. In these scenarios where there is a natural notion of groupings of observations, we demonstrate the interpretability of the learned features and how these structures of interaction correspond to physically meaningful systems. Furthermore, in such cases we show that the regularization employed by oi-VAE leads to better generalization and synthesis capabilities, especially in limited training data scenarios or when the training data might not fully capture the observed space of interest.
Researcher Affiliation Academia 1Paul G. Allen School of Computer Science and Engineering, University of Washington 2Institute for Learning & Brain Sciences and Department of Speech and Hearing Sciences, University of Washington.
Pseudocode Yes Algorithm 1 Collapsed VI for oi-VAE
Open Source Code No The paper does not provide a direct link or explicit statement about the availability of the open-source code for the methodology described.
Open Datasets Yes Using data collected from CMU s motion capture database we evaluated oi-VAE s ability to handle complex physical constraints and interactions across groups of joint angles while simultaneously identifying a sparse decomposition of human motion. We apply our oi-VAE method to infer low-rank representations of source-space MEG data where the groups are specified as the 40 regions defined in the HCP-MMP1 brain parcellation (Glasser et al., 2016).
Dataset Splits No The paper mentions training and testing on held-out examples but does not explicitly specify a separate 'validation' dataset split for hyperparameter tuning or early stopping, which is typically required for reproducibility of the validation phase.
Hardware Specification No The paper acknowledges "the support of NVIDIA Corporation for the donated GPU used for this research," but it does not specify the exact model or details of the GPU or any other hardware components like CPU, RAM, or specific cluster configurations.
Software Dependencies No The paper does not specify version numbers for any key software components or libraries used in the experiments (e.g., Python version, specific deep learning frameworks like TensorFlow or PyTorch, or scientific computing libraries).
Experiment Setup Yes In all experiments, we use λ = 1. To explore how the learned disentangled latent representation varies with latent dimension K, we use K = 4, 8, and 16. We applied oi-VAE with K = 20, λ = 10, and Alg. 1 for 10, 000 iterations.