Reproduce “Iterative Amortized Hierarchical VAE” 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
Simon W. Penninga, Ruud J. G. van Sloun
An Iterative Amortized Hierarchical VAE (IA-HVAE) improves the efficiency and accuracy of hierarchical VAEs for inverse problems. It achieves a 35x speed-up for iterative inference and produces superior reconstruction quality by combining amortized initialization with a novel linearly separable decoder.
Key insights
- Decomposing the decoder into linearly separable components in a transform domain allows for efficient, localized gradient propagation, circumventing the quadratic scaling issue of iterative refinement in deep HVAEs.
- A hybrid inference strategy, leveraging both amortized and iterative steps, effectively bridges the "amortization gap" in VAEs, yielding higher quality posterior estimates without full iterative cost.
- Integrating hierarchical latent structures with an efficient iterative refinement mechanism enables HVAEs to balance speed and accuracy, making them competitive for real-time inverse problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.