Research map / Probabilistic and scientific modeling / Variational and generative sampling
Variational Sparse Coding with Learned Thresholding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply shifted soft-thresholding to Laplacian samples | Variational sparse coding with a linear generator · Whitened image patches | Validation loss: 5.79E+02 -> 1.94E+02 | Helped |
| Apply shifted soft-thresholding to Gaussian samples | Variational sparse coding with a linear generator · Whitened image patches | Validation loss: 1.35E+03 -> 2.30E+02 | Helped |
| Apply shifted soft-thresholding to Gaussian samples | Variational sparse coding with a linear generator · Whitened image patches | Multi-Information: 7.36E+02 -> 1.76E+02 | Helped |
| Add a learned Gamma prior on the threshold instead of fixing the threshold (ablation) | Thresholded Gaussian · Whitened image patches | Validation loss: 2.30E+02 -> 2.70E+02 | Hurt |
| Add a learned Gamma prior on the threshold instead of fixing the threshold (ablation) | Thresholded Laplacian · Whitened image patches | Validation loss: 1.94E+02 -> 2.11E+02 | Hurt |
| Use a straight-through estimator rather than a subgradient (ablation) | Thresholded Laplacian · Collapse test, J=1 | Posterior collapse: NaN -> 4.17% | Helped |
| Use a straight-through estimator rather than a subgradient (ablation) | Thresholded Laplacian · Collapse test, J=20 | Feature collapse: 100.00% -> 0.00% | Helped |
| Use max-ELBO sampling instead of averaging sample losses (ablation) | Thresholded sparse variational inference · Fixed validation inputs | Support consistency (Jaccard index) | Helped |
| Use thresholded sparse priors rather than a Gaussian prior | VAE with a DNN generator · Fashion MNIST and CelebA | Multi-Information | Helped |
| Use Thresholded Gaussian or Thresholded Gaussian+Gamma rather than Spike-and-slab | VAE with a DNN generator · Fashion MNIST and CelebA | Performance as latent dimensionality increases | Helped |
| Increase the degree of sparsity using thresholded methods (ablation) | VAE with a DNN generator · Fashion MNIST and CelebA | Mean squared error | Helped |
Closest papers by meaning
- Be a Goldfish: Forgetting Bad Conditioning in Sparse Linear Regression via Variational Autoencoders (ICML 2025 · similarity 0.88)
- Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders (ICML 2025 · similarity 0.88)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.87)
- Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks (NeurIPS 2025 · similarity 0.87)
- Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth (ICML 2024 · similarity 0.86)
- Sparse Autoencoders, Again? (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.