Disentanglement via Latent Quantization

Authors: Kyle Hsu, William Dorrell, James Whittington, Jiajun Wu, Chelsea Finn

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the broad applicability of this approach by adding it to both basic data-reconstructing (vanilla autoencoder) and latent-reconstructing (Info GAN) generative models. For reliable evaluation, we also propose Info MEC, a new set of metrics for disentanglement that is cohesively grounded in information theory and fixes well-established shortcomings in previous metrics. Together with regularization, latent quantization dramatically improves the modularity and explicitness of learned representations on a representative suite of benchmark datasets. In particular, our quantized-latent autoencoder (QLAE) consistently outperforms strong methods from prior work in these key disentanglement properties without compromising data reconstruction.
Researcher Affiliation Academia Kyle Hsu Will Dorrell James C. R. Whittington Jiajun Wu Chelsea Finn Stanford University University College London Oxford University
Pseudocode Yes Algorithm 1 contains pseudocode for latent quantization and computing the quantization and commitment losses. Appendix A presents pseudocode for training a quantized-latent autoencoder (QLAE) in Algorithm 2 and a quantized-latent Info WGAN-GP [12, 1, 23] in Algorithm 3.
Open Source Code Yes Code for models and Info MEC metrics: https://github.com/kylehkhsu/latent_quantization.
Open Datasets Yes We benchmark on four established datasets: Shapes3D [9], MPI3D [20], Falcor3D [53], and Isaac3D [53].
Dataset Splits No The paper states 'We train on the entire dataset then evaluate on 10,000 i.i.d. samples' which describes a test set, and mentions hyperparameter tuning, but does not explicitly provide details on a validation dataset split.
Hardware Specification No The paper does not provide specific details on the hardware used, such as GPU or CPU models.
Software Dependencies No The paper acknowledges the use of several open-source software packages (e.g., NumPy, JAX, scikit-learn) but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We fix the number of latents in all methods to twice the number of sources. For quantized-latent models, we fix nv = 10 discrete values per codebook. We tune one key regularization hyperparameter per method per dataset with a thorough sweep (Table 11, Appendix C.2). Appendices C.2 and C.3 provide specific values for learning rates, batch sizes, update steps, loss weights, and network architectures.