Research map / Label-efficient and robust learning / Recognition and long-tail
FSPool: Learning Set Representations with Featurewise Sort Pooling
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 |
|---|---|---|---|
| Use FSPool with inverse sorting in place of an MLP set decoder and assignment-based loss | polygon auto-encoder · Rotating polygons, set sizes up to 128 | reconstruction error: 0.000 vs 1.272 (Chamfer loss, set size 16) | Helped |
| Replace the sum pooling readout with FSPool | CLEVR set encoder · CLEVR | accuracy: 99.05% ±0.17 -> 99.27% ±0.18 | Helped |
| Replace GIN's standard readout with FSPool | GIN · RDT-B | test accuracy: 84.8 ±1.7 -> 89.1 ±1.2 | Helped |
| Replace GIN's standard readout with FSPool | GIN · PTC | test accuracy: 59.9 ±2.4 -> 59.3 ±1.8 | Hurt |
| Replace GIN's standard readout with FSPool | GIN · graph classification datasets | epochs to best validation accuracy: FSPool takes fewer epochs on 6 of 9 datasets | Mixed |
| Replace sum pooling with FSPool | DSPN-RN set encoder · CLEVR bounding-box prediction | AP50: 87.2 ±3.0 -> 99.8 ±0.0 | Helped |
| Replace sum pooling with FSPool | DSPN-RN set encoder · CLEVR state prediction | AP∞: 30.2 ±9.2 -> 85.2 ±4.8 | Helped |
| Add a mask feature to distinguish present points from padding (ablation) | MNIST set auto-encoder · MNIST sets, σ=0.05 | test Chamfer loss (in 10 000ths): FSPool + FSUnpool 0.30 ±0.01; Sum + MLP 0.61 ±0.24 | Helped |
| Replace relaxed sorting with unrelaxed sorting at classification time (ablation) | FSPool model with frozen auto-encoder weights · Noisy MNIST classification | accuracy | No clear effect |
| Zero out the latent space (ablation) | FSPool polygon auto-encoder · Rotating polygons | ability to train | Hurt |
| Use a sum-pooling encoder and MLP decoder instead of FSPool and FSUnpool (ablation) | MNIST set auto-encoder · MNIST sets without mask feature | reconstruction quality: Chamfer loss: 0.44 ±0.01 vs 0.31 ±0.04 at σ=0.05 | Mixed |
| Set FSPool's piecewise-linear weight count to k=5 instead of k=20 (ablation) | GIN-FSPool · Graph classification datasets | test accuracy | Helped |
Closest papers by meaning
- Pooling by Sliced-Wasserstein Embedding (NeurIPS 2021 · similarity 0.87)
- Learning Representations of Sets through Optimized Permutations (ICLR 2019 · similarity 0.86)
- Stacking Deep Set Networks and Pooling by Quantiles (ICML 2024 · similarity 0.83)
- Polynomial Width is Sufficient for Set Representation with High-dimensional Features (ICLR 2024 · similarity 0.82)
- Cardinality-Aware Set Prediction and Top-$k$ Classification (NeurIPS 2024 · similarity 0.82)
- Exploring Diverse Representations for Open Set Recognition (AAAI 2024 · similarity 0.82)
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.