Research map / Architectures and efficiency / Pruning and efficient tuning
Learning Compatible Multi-Prize Subnetworks for Asymmetric Retrieval
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 PrunNet, a dense network with learnable per-weight scores that yields compatible multi-prize subnetworks at any capacity via post-training greedy pruning, trained with a conflict-aware gradient integration scheme for asymmetric image retrieval.
Problem
Asymmetric retrieval across platforms with different resource budgets requires compatible models at varying capacities; existing methods use pre-defined subnetworks and need new backward-compatible training for each new platform capacity.
Approach
Learn both scores and weights of a dense network so that subnetworks at any capacity can be obtained by greedy score-based pruning without retraining, and resolve gradient conflicts among dense and subnetwork losses via a conflict-aware gradient integration that projects and reweights gradients.
Key results
- On GLDv2/RParis/ROxford with ResNet-18, the dense network reaches 46.29 average mAP versus 45.41 for independent learning and 44.47 for SFSC.
- On VeRi-776, self-test at the full dense capacity reaches 67.82 mAP versus 66.11 for SFSC.
- A new 10% capacity subnetwork obtained without retraining reaches 44.67 average mAP on landmark benchmarks, above the BCT-based SFSC variant at 41.59.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train PrunNet with compatible multi-prize subnetworks | ResNet-18 retrieval models · GLDv2-test, RParis, ROxford (average) | mAP: 44.47 -> 46.29 | Helped |
| train PrunNet instead of independent learning | ResNet-18 retrieval models · In-shop | Recall@1: 86.14 -> 87.31 | Helped |
| train PrunNet instead of SFSC | ResNet-18 retrieval models · VeRi-776 | mAP: 66.11 -> 67.82 | Helped |
| generate a 10% capacity model by post-training pruning | PrunNet · GLDv2-test, RParis, ROxford (average), new capacity 10% | mAP: 43.13 -> 44.67 | Helped |
| optimize the score map rather than freezing it (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 43.00 -> 45.63 | Helped |
| use greedy pruning with inherited connections instead of separate score maps (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 42.23 -> 45.63 | Helped |
| use conflict-aware gradient integration instead of direct gradient summation (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 45.30 -> 45.63 | Helped |
| use conflict-aware gradient integration instead of direct loss combination (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 42.14 -> 45.63 | Helped |
| use conflict-aware gradient integration instead of Pareto integration (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 43.84 -> 45.63 | Helped |
| optimize weights as well as scores (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 2.99 -> 45.63 | Helped |
| train with more than one subnetwork (N = 4 rather than N = 1) (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 40.85 -> 45.63 | Helped |
| train with four subnetworks rather than two (ablation) | PrunNet · GLDv2-test, RParis, ROxford (average), φ20% self-test | mAP: 42.40 -> 45.63 | Helped |
Surprising findings
- Some prize subnetworks slightly outperform the dense network, which the authors attribute to removal of redundant weights.
- Freezing the scores or removing weight optimization degrades performance substantially, with the no-weight-optimization variant collapsing to near-zero mAP.
Limitations
- Pruning is unstructured, so resulting sparse subnetworks do not give direct acceleration without specialized hardware.
- Subnetwork capacities are pre-defined during training (N=4 by default), and new capacities rely on post-training pruning rather than explicit optimization.
Experimental setup
Datasets: GLDv2 In-shop VeRi-776 CIFAR-10
Benchmarks: GLDv2-test RParis ROxford In-shop VeRi-776
Metrics: mAP Recall@1
Baselines: Independent learning Joint learning O2O-SSPL BCT-S w/ SwitchNet Asymmetric-S w/ SwitchNet SFSC BCT-S w/ PrunNet Asymmetric-S w/ PrunNet
Builds on: SFSC edge-popup PCGrad BCT SSPL
Closest papers by meaning
- Adaptive Retrieval and Scalable Indexing for k-NN Search with Cross-Encoders (ICLR 2024 · similarity 0.84)
- USTAD: Unified Single-model Training Achieving Diverse Scores for Information Retrieval (ICML 2024 · similarity 0.84)
- Efficient stagewise pretraining via progressive subnetworks (ICLR 2025 · similarity 0.83)
- Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval (NeurIPS 2024 · similarity 0.83)
- Hierarchical Retrieval: The Geometry and a Pretrain-Finetune Recipe (NeurIPS 2025 · similarity 0.83)
- Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval (AAAI 2025 · similarity 0.83)
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.