Reproduce “Adaptive Compression of the Latent Space in Variational Autoencoders” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gabriela Sejnova, Michal Vavrecka, Karla Stepanova
Researchers at the Czech Institute of Informatics, Robotics and Cybernetics at CTU in Prague developed ALD-VAE, an adaptive method that automatically determines the optimal latent space dimensionality in Variational Autoencoders during a single training run. This approach achieves comparable model performance to conventionally tuned VAEs with substantially reduced computational overhead, streamlining VAE deployment.
Key insights
- Dynamically pruning latent space neurons based on a multi-metric stopping criterion can effectively identify a balanced optimum for VAE performance.
- Monitoring the slopes of key performance metrics (FID, reconstruction loss, Silhouette score) provides a robust signal for adaptive latent space compression.
- The method demonstrates robustness to varying initial latent space sizes and random seeds, consistently converging to competitive performance levels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.