Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Representations of Sets through Optimized Permutations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add PO-U permutation learning and an LSTM set representation to BAN | BAN · VQA v2 validation | Overall accuracy: 65.96 ± 0.16 -> 66.33 ± 0.09 | Helped |
| Add an LSTM set representation without permutation learning | BAN · VQA v2 validation | Overall accuracy: 65.96 ± 0.16 -> 66.06 ± 0.13 | No clear effect |
| Add PO-U permutation learning to the BAN + LSTM set representation | BAN + LSTM · VQA v2 validation | Overall accuracy: 66.06 ± 0.13 -> 66.33 ± 0.09 | Helped |
| Use linear rather than uniform assignment initialization (ablation) | PO model · Image mosaics | mean squared error: MNIST 5 × 5: 0.45 -> 0.01; CIFAR10 5 × 5: 1.26 -> 0.70; ImageNet 64 × 64 2 × 2: 0.14 -> 0.16 | Mixed |
| Use linear rather than uniform assignment initialization (ablation) | PO model · Implicit image classification | accuracy: MNIST 5 × 5: 69.2 -> 89.8; ImageNet 64 × 64 2 × 2: 24.6 -> 23.4 | Mixed |
| Use pairwise permutation optimization instead of LinAssign | Image mosaic reconstruction · CIFAR10 2 × 2 | mean squared error: 0.37 -> 0.11 | Helped |
| Use pairwise permutation optimization instead of LinAssign | Image mosaic reconstruction · MNIST 5 × 5 | mean squared error: LinAssign 0.08; PO-U 0.45; PO-LA 0.01 | Mixed |
| Use pairwise permutation optimization instead of LinAssign | Implicit image classification · ImageNet 64 × 64 2 × 2 | accuracy: 13.1 -> 24.6 | Helped |
| Use pairwise permutation optimization instead of LinAssign | Implicit image classification · CIFAR10 5 × 5 | accuracy: 32.5 -> 29.7 | Hurt |
| Use pairwise comparisons that model relative tile relations rather than elementwise absolute assignments | Set-to-sequence or grid permutation · ImageNet 64 × 64 implicit image classification | accuracy: 13.1 -> 24.6 | Helped |
| Rely on uniform-initialized pairwise costs for identical blank tiles | PO-U · MNIST image mosaics, higher tile counts | mean squared error: 0.08 -> 0.45 at 5 × 5, versus LinAssign | Hurt |
Closest papers by meaning
- Learning Distributions over Permutations and Rankings with Factorized Representations (ICLR 2026 · similarity 0.88)
- Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks (ICML 2019 · similarity 0.86)
- The Weighted Kendall and High-order Kernels for Permutations (ICML 2018 · similarity 0.86)
- DeepPermNet: Visual Permutation Learning (CVPR 2017 · similarity 0.86)
- FSPool: Learning Set Representations with Featurewise Sort Pooling (ICLR 2020 · similarity 0.86)
- REVISITING MULTI-PERMUTATION EQUIVARIANCE THROUGH THE LENS OF IRREDUCIBLE REPRESENTATIONS (ICLR 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.