Research map / Reinforcement learning / Policy gradient and actor-critic
Weighted Sampling without Replacement for Deep Top-$k$ Classification
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 WSWR top-k loss instead of cross-entropy | ResNet-18 on CIFAR-100 with label noise · CIFAR-100 noise 0.5 | Top-5 accuracy: 88.047 vs 81.207 | Helped |
| Use WSWR top-k loss instead of cross-entropy | ResNet-18 on CIFAR-100 with label noise · CIFAR-100 noise 0.0 | Top-5 accuracy: 94.043 vs 93.417 | Helped |
| Increase number of sampled trajectories m from 1 to 10000 (ablation) | WSWR on CIFAR-100 noise 0.5 · CIFAR-100 noise 0.5 | Top-5 accuracy: 87.733 (m1) to 88.424 (m10000) | Helped |
| Increase number of sampled trajectories m from 100 to 10000 (ablation) | WSWR on CIFAR-100 noise 0.5 · CIFAR-100 noise 0.5 | Training speed (batch/s): 19.713 (m100) to 0.744 (m10000) | Hurt |
| Use WSWR top-k loss instead of cross-entropy | ResNet-18 on CIFAR-100 with label noise · CIFAR-100 noise 0.5 | Top-1 accuracy: 60.020 vs 48.713 | Helped |
| Use WSWR top-k loss instead of cross-entropy at k=50 | ResNet-18 on CIFAR-100 noise 1.0 · CIFAR-100 top-50 (noise 1.0) | Top-50 accuracy: 11.237 vs 15.980 (CE is higher) | Helped |
| Use WSWR top-k loss instead of cross-entropy on ImageNet subsets | ResNet-50 trained from scratch on ImageNet-1K subsets · ImageNet-1K 5% subset | Top-5 accuracy: 64.12 vs 58.56 (m1000) | Helped |
| Use WSWR top-k loss instead of cross-entropy on full ImageNet-1K | ResNet-50 trained from scratch on ImageNet-1K 100% · ImageNet-1K 100% | Top-1 accuracy: 69.59 vs 75.76 (CE) | Hurt |
| Increase number of sampled trajectories m from 100 to 1000 (ablation) | WSWR on ImageNet-1K subsets · ImageNet-1K 10% subset | Top-5 accuracy: 76.87 (m1000) vs 73.91 (m100) on 10%; 89.20 vs 88.12 on 25% | Mixed |
| Increase number of sampled trajectories m from 100 to 1000 (ablation) | WSWR on CIFAR-100 noise 0.5 · CIFAR-100 noise 0.5 | Top-5 accuracy: 88.047 (m100) vs 88.247 (m1000) | No clear effect |
| Use sort-based relaxed top-k losses (RBitonic, RNeuralSort, RSoftSort) instead of WSWR at k=50 | ResNet-18 on CIFAR-100 noise 1.0 · CIFAR-100 top-50 (noise 1.0) | Top-50 accuracy: e.g., RNeuralSort 0.025 vs WSWR 11.237 | Hurt |
Closest papers by meaning
- ST$_k$: A Scalable Module for Solving Top-k Problems (NeurIPS 2024 · similarity 0.85)
- Provably Reliable Classifier Guidance via Cross-Entropy Control (NeurIPS 2026 · similarity 0.84)
- Decoupled Kullback-Leibler Divergence Loss (NeurIPS 2024 · similarity 0.84)
- Foundations of Top-$k$ Decoding for Language Models (NeurIPS 2025 · similarity 0.84)
- Rethinking Losses for Diffusion Bridge Samplers (NeurIPS 2025 · similarity 0.84)
- Set Learning for Accurate and Calibrated Models (ICLR 2024 · 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.