Research map / Architectures and efficiency / Attention and transformers
Ensembling Pruned Attention Heads For Uncertainty-Aware Efficient Transformers
Firas Gabetni, Giuseppe Curci, Andrea Pilzer, Subhankar Roy, Elisa Ricci, Gianni Franchi
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 Hydra Ens (Circ) instead of Deep Ensembles | Hydra Ensembles · ImageNet-1K OOD avg | AUROC: 85.48 -> 86.29 | Improved |
| Merge three Taylor-pruned members into Hydra Ens (ablation) | Taylor pruning · ImageNet-1K | accuracy: 80.68 -> 81.20 | Improved |
| Add ensembling of three circuit-pruned members to CircAvg (ablation) | CircAvg · SST-2 OOD avg | AUROC: 75.04 -> 77.6 | Improved |
| Add ensembling of three pruned members to CircAvg (ablation) | CircAvg · Zero-shot classification avg | accuracy: 71.91 -> 74.00 | Improved |
| Use circuit-based head selection instead of Taylor pruning, zero-shot (ablation) | OpenCLIP ViT-B/32 pruned heads · Zero-shot classification avg | accuracy: 60.75 -> 71.91 | Improved |
| Zero-shot Taylor head pruning of the dense model (ablation) | ViT-B/16 · ImageNet-1k zero-shot pruning | accuracy: 80.67 -> 66.46 | Hurt |
| Use circuit-based instead of Taylor head pruning, zero-shot (ablation) | ViT-B/16 · ImageNet-1k zero-shot pruning | accuracy: 66.46 -> 73.61 | Improved |
| Use circuit pruning (CircOOD) instead of dense model for ID-OOD separation (ablation) | ViT-B/16 · CIFAR-100 OOD CLS-token distance | Euclidean distance: 1.75 -> 2.26 | Improved |
| Extract the circuit for accuracy (CircAcc) instead of OOD (CircOOD) (ablation) | ViT-B/16 · ImageNet-1K | accuracy: 71.11 -> 73.41 | Improved |
| Extract the circuit for OOD (CircOOD) instead of accuracy (CircAcc) (ablation) | ViT-B/16 · ImageNet-1K OOD avg | AUROC: 80.80 -> 81.89 | Improved |
| Use Hydra Ens (Circ) instead of ViLU | Hydra Ensembles · Zero-shot ImageNet-1k OOD | AUROC: +1.3% AUROC, -3.5% FPR95, +4.0% AUPR | Improved |
| Use Hydra Ensembles instead of Deep Ensembles on SST-2 | Hydra Ensembles · SST-2 OOD avg | AUROC: AUROC (+2.8%), FPR95 (-7.6%), AUPR (+2.2%) | Improved |
Closest papers by meaning
- Hyperparameter Ensembles for Robustness and Uncertainty Quantification (NeurIPS 2020 · similarity 0.87)
- Neural Ensemble Search for Uncertainty Estimation and Dataset Shift (NeurIPS 2021 · similarity 0.86)
- FiLM-Ensemble: Probabilistic Deep Learning via Feature-wise Linear Modulation (NeurIPS 2022 · similarity 0.86)
- Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction (ICLR 2025 · similarity 0.84)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.84)
- Asymmetric Duos: Sidekicks Improve Uncertainty (NeurIPS 2025 · 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.
