Research map / Architectures and efficiency / CNN design
Adaptative Inference Cost With Convolutional Neural Mixture Models
Adria Ruiz, Jakob Verbeek
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 |
|---|---|---|---|
| use expectation propagation at inference instead of 1 Monte-Carlo sample (ablation) | CNMM · CIFAR-100 | top-1 accuracy: 71.2 -> 74.4 | Improved |
| use 30 Monte-Carlo samples instead of expectation propagation (ablation) | CNMM · CIFAR-100 | top-1 accuracy: 74.4 -> 74.6 | No clear effect |
| use 5 Monte-Carlo samples instead of expectation propagation (ablation) | CNMM · CIFAR-100 | top-1 accuracy: 74.4 -> 74.4 | No clear effect |
| add stochastic sampling of mixture variables during training (vs training with expectations) (ablation) | CNMM · CIFAR-100 | accuracy under pruning | Improved |
| add stochastic mixture over networks (vs deterministic summation with early exits) (ablation) | CNMM · CIFAR-100 | accuracy vs FLOPs | Improved |
| add intermediate classifiers with per-depth losses (ablation) | CNMM · CIFAR-10, CIFAR-100, Cityscapes | accuracy at low FLOPs | Improved |
| add intermediate classifiers, final-step accuracy (ablation) | CNMM · CIFAR-10, CIFAR-100, Cityscapes | final accuracy | Mixed |
| prune the trained CNMM by removing low-probability networks | CNMM · CIFAR-10/100 | accuracy vs FLOPs: factor two FLOP reduction | No clear effect |
| prune the trained CNMM by removing low-probability networks | CNMM · Cityscapes val | mIoU: about 25% pruning | No clear effect |
| CNMM instead of ESPNet | CNMM · Cityscapes val | mIoU: +0.5 points mIoU, FLOPs reduced by 45% | Improved |
Closest papers by meaning
- VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024 · similarity 0.84)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.84)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (CVPR 2025 · similarity 0.84)
- Bayesian Test-Time Adaptation for Vision-Language Models (CVPR 2025 · similarity 0.84)
- Toward Efficient Inference for Mixture of Experts (NeurIPS 2024 · similarity 0.84)
- Vision Transformers as Probabilistic Expansion from Learngene (ICML 2024 · 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.
