Research map / Label-efficient and robust learning / Recognition and long-tail
ProtoAttend: Attention-Based Prototypical Learning
Sercan O. Arik, Tomas Pfister
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 sparsemax instead of softmax attention to reduce prototypes (ablation) | ProtoAttend · MNIST | prototypes for 50% of decision: 365 -> 2 | Improved |
| use sparsemax instead of softmax attention (ablation) | ProtoAttend · DBPedia | accuracy (%): 98.20 -> 97.74 | Hurt |
| add sparsity regularization on top of sparsemax (ablation) | ProtoAttend · Fashion-MNIST | prototypes for 50% of decision: 4 -> 1 | Improved |
| add sparsity regularization on top of sparsemax (ablation) | ProtoAttend · CIFAR-10 | accuracy (%): 91.44 -> 91.26 | Hurt |
| use ProtoAttend with sparsemax + sparsity reg. instead of baseline encoder | baseline encoder · CIFAR-10 label noise 0.8 | test accuracy (%): 57.02 -> 60.50 | Improved |
| use ProtoAttend instead of baseline encoder | baseline encoder · CIFAR-10 label noise 0.4 | test accuracy (%): 77.47 -> 80.04 | Improved |
| use ProtoAttend instead of dropout | dropout · CIFAR-10 label noise 0.8 | test accuracy (%): 56.76 -> 60.50 | Improved |
| train only with prototype output loss L(y,yhat(1)) instead of adding yhat(0) loss (ablation) | ProtoAttend training loss · Fashion-MNIST | accuracy (%) of yhat(0): 94.37 -> 10.92 | Hurt |
| train only with L(y,yhat(0)) instead of adding yhat(1) loss (ablation) | ProtoAttend training loss · Fashion-MNIST | accuracy (%) of yhat(1): 94.38 -> 13.13 | Hurt |
| use yhat(0.5) loss only instead of L0+L1 (ablation) | ProtoAttend training loss · Fashion-MNIST | confidence for correct cases: 0.931 -> 0.927 | Hurt |
| add L(y,yhat(0.5)) to L0+L1 (ablation) | ProtoAttend training loss · Fashion-MNIST | accuracy (%) of yhat(1): 94.38 -> 94.45 | No clear effect |
| use confidence score from prototype label agreement | ProtoAttend · confidence-controlled prediction | accuracy vs ratio of samples | Improved |
Closest papers by meaning
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.86)
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.85)
- A Robust Prototype-Based Network with Interpretable RBF Classifier Foundations (AAAI 2025 · similarity 0.85)
- ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.85)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.85)
- Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time (CVPR 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.
