Research map / Learning theory and algorithms / Optimal transport
LaCoOT: Layer Collapse through Optimal Transport
Victor Quétu, Zhu Liao, Nour Hezbri, Fabio Pizzati, Enzo Tartaglione
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| raise regularization weight lambda from 0.01 to 5 (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Top-1 accuracy after block removal: 73.04 -> 90.99 | Improved |
| add the Max-Sliced Wasserstein regularization (lambda>0) before ranking blocks (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Top-1 vs MACs | Improved |
| use LaCoOT importance score instead of block influence or random removal (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Performance/compression trade-off | Improved |
| use Sliced Wasserstein instead of Max-Sliced Wasserstein (ablation) | LaCoOT (lambda=5) · ResNet-18 on CIFAR-10 | Top-1 vs MACs | Hurt |
| reduce number of projections to 1 (ablation) | LaCoOT (lambda=5) · ResNet-18 on CIFAR-10 | Top-1 vs MACs | Hurt |
| reduce batch size used for the distance (ablation) | LaCoOT (lambda=5) · ResNet-18 on CIFAR-10 | Top-1 | Hurt |
| replace MSWD with KL divergence (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Top-1 vs MACs | Hurt |
| replace MSWD with l1, l2 or MMD (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Top-1 vs MACs | No clear effect |
| add LaCoOT regularized finetuning to DiT-XL/2 with two blocks removed (ablation) | DiT-XL/2 on ImageNet · ImageNet generation | FID-50k: FID-50k twice as low | Improved |
| add DepGraph (0.2) on top of LaCoOT (ablation) | LaCoOT (lambda=5) · ResNet-18 on CIFAR-10 | MACs / top-1: 64.69 -> 57.22 MACs, 90.99 -> 90.96 top-1 | Improved |
| add DepGraph (0.3) on top of LaCoOT (ablation) | LaCoOT (lambda=5) · ResNet-18 on CIFAR-10 | Top-1: 90.99 -> 89.27 | Hurt |
| add regularization for 3x3 conv replaced by avgpool and 1x1 conv (lambda 0 to 0.1) (ablation) | LaCoOT · ResNet-18 on CIFAR-10 | Performance loss vs full model: 0.98% -> 0.18% loss | Improved |
Closest papers by meaning
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.87)
- MinMax Methods for Optimal Transport and Beyond: Regularization, Approximation and Numerics (NeurIPS 2020 · similarity 0.86)
- Expectile Regularization for Fast and Accurate Training of Neural Optimal Transport (NeurIPS 2024 · similarity 0.86)
- Achieving Robustness in Classification Using Optimal Transport With Hinge Regularization (CVPR 2021 · similarity 0.86)
- Revealing the Structure of Deep Neural Networks via Convex Duality (ICML 2021 · similarity 0.85)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.85)
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.
