Research map / Probabilistic and scientific modeling / Variational and generative sampling
Invertible DenseNets with Concatenated LipSwish
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 |
|---|---|---|---|
| use densely connected blocks (i-DenseNet) instead of residual blocks with similar parameters | Residual Flow · CIFAR10 | bits per dimension: 3.28 -> 3.25 | Helped |
| use densely connected blocks instead of residual blocks | Residual Flow · ImageNet32 | bits per dimension: 4.01 -> 3.98 | Helped |
| add learnable weighted concatenation (ablation) | i-DenseNets · toy data 2 circles | negative log-likelihood (nats): 3.32 -> 3.30 | Helped |
| use CLipSwish activation instead of LipSwish (ablation) | i-DenseNet · CIFAR10 small models | bits per dimension: 3.39 -> 3.37 | Helped |
| use CLipSwish activation instead of LipSwish (ablation) | Residual Flow · CIFAR10 small models | bits per dimension: 3.42 -> 3.38 | Helped |
| use LeakyLSwish instead of LipSwish (ablation) | i-DenseNet · CIFAR10 small models | bits per dimension: 3.39 -> 3.39 | No clear effect |
| increase concatenation depth from 3 to 5 at equal parameters (ablation) | i-DenseNet · CIFAR10 small models | bits per dimension: 3.37 -> 3.42 | Hurt |
| enlarge ResNet connections in Residual Flow to match i-DenseNet size instead of dense connections (ablation) | Residual Flow · CIFAR10 small models | bits per dimension: 3.42 -> 7.02 | Hurt |
| dense blocks instead of residual blocks in hybrid model, lambda=1 | hybrid flow on CIFAR10 · CIFAR10 hybrid | accuracy: 70.32% -> 75.67% | Helped |
| dense blocks instead of residual blocks in hybrid model, lambda=0 | hybrid flow on CIFAR10 · CIFAR10 hybrid | accuracy: 91.78% -> 92.40% | Helped |
Closest papers by meaning
- Residual Flows for Invertible Generative Modeling (NeurIPS 2019 · similarity 0.91)
- Approximation theory for 1-Lipschitz ResNets (NeurIPS 2025 · similarity 0.88)
- Invertible Residual Networks (ICML 2019 · similarity 0.88)
- Monotone, Bi-Lipschitz, and Polyak-Łojasiewicz Networks (ICML 2024 · similarity 0.86)
- A Rescaling-Invariant Lipschitz Bound Based on Path-Metrics for Modern ReLU Network Parameterizations (ICML 2025 · similarity 0.85)
- CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices (NeurIPS 2025 · 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.