Research map / Architectures and efficiency / Pruning and efficient tuning
Partial Fusion of Neural Networks: Efficient Tradeoffs Between Ensembles and Weight Aggregation
Fabian Morelli, Stephan Eckstein
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 alpha from 0 to 1 (fewer to no isolated neurons to full ensemble) before fine-tuning (ablation) | Partial OT Fusion MLP · Split MNIST MLP, fused (no fine-tuning) | accuracy (%): 84.1 -> 88.1 | Improved |
| raise alpha from 0 to 1 before fine-tuning (ablation) | Partial OT Fusion ResNet-18 · Split CIFAR-10 ResNet-18, fused (no fine-tuning) | accuracy (%): 66.4 -> 91.3 | Improved |
| fine-tune the fused model on 1% data (alpha=0) | Partial OT Fusion MLP · Split MNIST, fine-tuned with 1% data | accuracy (%): 84.1 -> 95.1 | Improved |
| fine-tune the fused model on 5% data (alpha=0) | Partial OT Fusion ResNet-18 · Split CIFAR-10, fine-tuned with 5% data | accuracy (%): 66.4 -> 85.3 | Improved |
| raise alpha from 0 to 1 after fine-tuning (ablation) | Partial OT Fusion ResNet-18 · Split CIFAR-10, fine-tuned with 5% data | accuracy (%): 85.3 -> 91.8 | Improved |
| raise alpha from 0 to 1 after fine-tuning (ablation) | Partial OT Fusion MLP · Split MNIST, fine-tuned with 1% data | accuracy (%): 95.1 -> 96.5 | Improved |
| use partial fusion (alpha=1.0) instead of random widening of Model A (ablation) | fine-tuned wider network · Split CIFAR-10, fine-tuned with 5% data | accuracy (%): 49.0 -> 91.7 | Improved |
| use partial fusion instead of Net2WiderNet widening, alpha=1.0 (ablation) | fine-tuned wider network · Split CIFAR-10, fine-tuned with 5% data | accuracy (%): 86.6 -> 91.7 | Improved |
| use partial fusion instead of Net2WiderNet widening, alpha=1.0 (ablation) | fine-tuned wider network · Split MNIST, fine-tuned with 1% data | accuracy (%): 95.0 -> 96.5 | Improved |
| increase alpha from 0 to 0.4 at width 25 (ablation) | Partial OT Fusion MLP · MNIST MLP same data, hidden width 25 | accuracy (%): 64.7 -> 85.0 | Improved |
| increase alpha from 0 to 1 at width 400 (ablation) | Partial OT Fusion MLP · MNIST MLP same data, hidden width 400 | accuracy (%): 97.2 -> 97.9 | No clear effect |
| use fixed-point iteration instead of greedy weight-based or activation-based matching (ablation) | Partial OT Fusion · Split MNIST MLP | accuracy | Improved |
Closest papers by meaning
- The Diversified Ensemble Neural Network (NeurIPS 2020 · similarity 0.88)
- $C^2M^3$: Cycle-Consistent Multi-Model Merging (NeurIPS 2024 · similarity 0.87)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.86)
- Expand-and-Cluster: Parameter Recovery of Neural Networks (ICML 2024 · similarity 0.86)
- Expressivity of Neural Networks with Random Weights and Learned Biases (ICLR 2025 · similarity 0.85)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (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.
