Research map / Architectures and efficiency / Pruning and efficient tuning
MDP: Multidimensional Vision Model Pruning with Latency Constraint
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MDP, a multidimensional structural pruning method that jointly prunes channels, transformer dimensions, and blocks under a latency budget by solving a MINLP with a decomposed multi-dimensional latency model.
Problem
Structural pruning is limited to fine-grained granularities and relies on linear latency models that fail for transformers, so aggressive high-ratio pruning under strict latency budgets is suboptimal.
Approach
Encode all prunable dimensions (channels, block removal, and for transformers embedding, heads, Q/K, V, MLP) with one-hot and binary variables, build a lookup-table-based latency constraint that captures interactions across dimensions, and solve the resulting MINLP with Pyomo/MindtPy to find the globally optimal pruned structure under a latency budget.
Key results
- On ImageNet ResNet50 at 85% pruning, MDP reaches 70.0 Top-1 at 5.21x speedup vs HALP's 68.1 at 3.90x.
- On DEIT-Base, MDP reaches 78.21 Top-1 at 5.57x speedup vs Isomorphic's 77.50 at 5.20x.
- On nuScenes with StreamPETR at 45% pruning, MDP gets 0.451 mAP and 1.18x speedup vs dense 0.449 mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MDP-85% instead of HALP-85% | ResNet50 · ImageNet | Top-1 accuracy (%): 68.1 -> 70.0 | Helped |
| Apply MDP-85% instead of HALP-85% | ResNet50 · ImageNet | speedup: 3.90 -> 5.21 | Helped |
| Apply MDP-19% instead of Isomorphic-T | DeiT-Base · ImageNet | Top-1 accuracy (%): 77.50 -> 78.21 | Helped |
| Apply MDP-19% instead of Isomorphic-T | DeiT-Base · ImageNet | speedup: 5.20 -> 5.57 | Helped |
| Prune StreamPETR with MDP-45% | StreamPETR · NuScenes | mAP: 0.449 -> 0.451 | Helped |
| Prune StreamPETR with MDP-45% | StreamPETR · NuScenes | speedup: 1.00 -> 1.18 | Helped |
| Add multi-granularity pruning (block pruning) to OCP+LLM (ablation) | ResNet50 MDP · ImageNet | accuracy-latency tradeoff | Helped |
| Add multi-dimensional latency modeling to OCP+LLM (ablation) | ResNet50 MDP · ImageNet | accuracy-latency tradeoff | Helped |
| Add multi-dimensional latency modeling to MGP+LLM (ablation) | ResNet50 MDP · ImageNet | accuracy-latency tradeoff | Helped |
| Add multi-granularity pruning to OCP+MDLM (ablation) | ResNet50 MDP · ImageNet | accuracy-latency tradeoff | Helped |
Limitations
- Relies on a pre-computed latency lookup table on target hardware.
- Block-level and multi-dimension decomposition is model-specific (CNN vs transformer decomposition differs).
Experimental setup
Datasets: ImageNet Pascal VOC nuScenes
Benchmarks: ImageNet Pascal VOC nuScenes
Metrics: Top-1 accuracy Top-5 accuracy speedup mAP NDS FLOPs
Baselines: HALP SMCP NViT Isomorphic LayerPrune EagleEye DBP StreamPETR
Builds on: HALP NViT Taylor channel importance Pyomo
Closest papers by meaning
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.88)
- UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 2024 · similarity 0.88)
- CORP: Closed-Form One-shot Representation-Preserving Structured Pruning for Transformers (NeurIPS 2026 · similarity 0.88)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.88)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.88)
- Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · similarity 0.87)
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.