Local Disentanglement in Variational Auto-Encoders Using Jacobian $L_1$ Regularization
Authors: Travers Rhodes, Daniel Lee
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our novel JL1-VAE framework to a variety of datasets, giving qualitative and quantitative results showing that our added L1 cost can encourage local alignment of the axes of the latent representation with individual factors of variation. |
| Researcher Affiliation | Academia | Travers Rhodes Department of Computer Science Cornell Tech, Cornell University New York, NY 10044 tsr42@cornell.edu Daniel D. Lee Department of Electrical and Computer Engineering Cornell Tech, Cornell University New York, NY 10044 ddl46@cornell.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using `disentanglement_lib` which is open-source, but it does not state that the authors are releasing their own code for the JL1-VAE methodology. |
| Open Datasets | Yes | The first are natural images in grayscale taken by Olshausen and Field [7]... The second is a dataset of simulated 64x64-pixel grayscale images of three black dots on a white background, inspired by [34]. ... Finally, we also apply our approach to tiled images of a real robotic arm taken from the MPI3D-real dataset [35], licensed under Creative Commons Attribution 4.0 International License. |
| Dataset Splits | No | The paper describes training on batches and total batches, but does not specify explicit train/validation/test dataset splits with percentages, counts, or references to predefined splits. |
| Hardware Specification | Yes | We train these models on a Nvidia Quadro V100 hosted locally and one hosted on Google Cloud. ... we train on a Nvidia Quadro V100s hosted locally. |
| Software Dependencies | No | The paper mentions software like Adam optimizer and `disentanglement_lib` but does not provide specific version numbers for these or other key software components, which is required for reproducibility. |
| Experiment Setup | Yes | We use the Adam optimizer with a learning rate of 0.0001 (matching [11]). We use linear annealing from 0 to the final hyperparameter value over the first 100,000 batches for both the beta hyperparameter and JL1-VAE s γ parameter in our implementations for JL1-VAE and β-VAE (unlike [11]). ... We use a latent dimension of ten for all experiments |