Research map / Architectures and efficiency / RNNs and normalization
Implicit Rank-Minimizing Autoencoder
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 |
|---|---|---|---|
| insert extra linear layers between encoder and decoder | deterministic AE · CelebA FID, Gaussian sampling | FID: 52.8 -> 42.5 | Helped |
| insert extra linear layers between encoder and decoder | deterministic AE · MNIST FID, Gaussian sampling | FID: 55.0 -> 37.4 | Helped |
| insert extra linear layers between encoder and decoder | deterministic AE · MNIST classification, 10 labels | error rate (%): 31.4 -> 12.0 | Helped |
| insert extra linear layers between encoder and decoder | deterministic AE · MNIST classification, 60000 labels | error rate (%): 1.9 -> 1.9 | No clear effect |
| fix the inserted linear matrices (no training) (ablation) | IRMAE · MNIST sampling quality and latent rank | qualitative / rank | Hurt |
| add nonlinearity between the linear matrices (ablation) | IRMAE · generation quality and latent rank | qualitative / rank | Hurt |
| share weights across inserted linear matrices (ablation) | IRMAE · generation quality and latent rank | qualitative / rank | Hurt |
| increase linear depth from 2 to 4 layers (ablation) | IRMAE · MNIST FID | FID: 44.0 -> 30.1 | Helped |
| increase linear depth from 8 to 12 layers (ablation) | IRMAE · MNIST FID | FID: 37.4 -> 62.6 | Hurt |
| increase initial variance of linear layers from 1x to 4x (ablation) | IRMAE · MNIST FID | FID: 37.4 -> 49.0 | Hurt |
| use IRMAE instead of AE at latent dim 512 | CelebA autoencoder · CelebA FID | FID: 53.9 -> 42.0 | Helped |
| use IRMAE instead of AE at latent dim 32 | CelebA autoencoder · CelebA FID | FID: 78.2 -> 81.6 | Hurt |
Closest papers by meaning
- Gradient Origin Networks (ICLR 2021 · similarity 0.86)
- Variational Lossy Autoencoder (ICLR 2017 · similarity 0.85)
- Structure by Architecture: Structured Representations without Regularization (ICLR 2023 · similarity 0.85)
- Improving Generalization and Convergence by Enhancing Implicit Regularization (NeurIPS 2024 · similarity 0.84)
- A Statistical Analysis of Wasserstein Autoencoders for Intrinsically Low-dimensional Data (ICLR 2024 · similarity 0.84)
- Improved Baselines with Representation Autoencoders (NeurIPS 2026 · similarity 0.84)
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.