Research map / Label-efficient and robust learning / Semi-supervised learning
Deconstructing the Ladder Network Architecture
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 |
|---|---|---|---|
| Add lateral connections to the encoder-decoder architecture (ablation) | Ladder Network · PI MNIST, semi-supervised (100 and 1000 labels) | AER (%): 100 labels: 1.086 vs 16.390; 1000 labels: 1.017 vs 5.251 | Helped |
| Inject noise at every layer rather than only the first layer (ablation) | Ladder Network · PI MNIST, 100 and 1000 labels | AER (%): 100 labels: 1.856 -> 1.086; 1000 labels: 1.381 -> 1.017 | Helped |
| Retain reconstruction penalties beyond the first layer (ablation) | Ladder Network · PI MNIST, 100 labels | AER (%): 1.691 -> 1.086 | Helped |
| Retain reconstruction penalties beyond the first layer (ablation) | Ladder Network · PI MNIST, 1000 and 60000 labels | AER (%): 1000 labels: 1.053 -> 1.017; 60000 labels: 0.608 -> 0.608 | No clear effect |
| Keep noise injection and reconstruction at layers beyond the first (ablation) | Ladder Network · PI MNIST, 100 and 60000 labels | AER (%): 100 labels: 1.856 -> 1.086; 60000 labels: 0.732 -> 0.608 | Helped |
| Keep noise injection and reconstruction at layers beyond the first (ablation) | Ladder Network · PI MNIST, 1000 labels | AER (%): 1.058 -> 1.017 | No clear effect |
| Retain sigmoid nonlinearity in the vanilla combinator (ablation) | Ladder Network · PI MNIST, all label counts | AER (%): 100 labels: 1.608 -> 1.086; 1000 labels: 1.223 -> 1.017; 60000 labels: 0.633 -> 0.608 | Helped |
| Retain the multiplicative term in the vanilla combinator (ablation) | Ladder Network · PI MNIST, all label counts | AER (%): 100 labels: 3.041 -> 1.086; 1000 labels: 1.735 -> 1.017; 60000 labels: 0.674 -> 0.608 | Helped |
| Use a nonlinear combinator rather than simple linear addition (ablation) | Ladder Network · PI MNIST, all label counts | AER (%): 100 labels: 5.027 -> 1.086; 1000 labels: 2.769 -> 1.017; 60000 labels: 0.849 -> 0.608 | Helped |
| Use the vanilla lateral-first initialization rather than random initialization (ablation) | Ladder Network · PI MNIST, all label counts | AER (%): 100 labels: 1.232 -> 1.086; 1000 labels: 1.011 -> 1.017; 60000 labels: 0.614 -> 0.608 | Helped |
| Use the vanilla lateral-first initialization rather than reverse initialization (ablation) | Ladder Network · PI MNIST, all label counts | AER (%): 100 labels: 1.305 -> 1.086; 1000 labels: 1.031 -> 1.017; 60000 labels: 0.631 -> 0.608 | Helped |
| Replace the vanilla combinator with an unconstrained Gaussian combinator | Ladder Network · PI MNIST, 100 and 1000 labels | AER (%): 100 labels: 1.086 -> 1.064; 1000 labels: 1.017 -> 0.983 | Helped |
Closest papers by meaning
- Exact Combinatorial Multi-Class Graph Cuts for Semi-Supervised Learning (AAAI 2026 · similarity 0.87)
- Meta Propagation Networks for Graph Few-shot Semi-supervised Learning (AAAI 2022 · similarity 0.86)
- Big Self-Supervised Models are Strong Semi-Supervised Learners (NeurIPS 2020 · similarity 0.86)
- Recurrent Ladder Networks (NeurIPS 2017 · similarity 0.85)
- Continuous Partitioning for Graph-Based Semi-Supervised Learning (NeurIPS 2024 · similarity 0.84)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 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.