Research map / Label-efficient and robust learning / Recognition and long-tail
Differentiable Top-k Classification Learning
Felix Petersen, Hilde Kuehne, Christian Borgelt, Oliver Deussen
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 |
|---|---|---|---|
| replace a single top-1 objective with a mixture of top-1 and top-5 objectives | ResNet18 with DiffSortNets · CIFAR-100 | Top-1 accuracy: 61.27 -> 62.00 | Improved |
| replace a single top-1 objective with a mixture of top-1 and top-5 objectives | ResNet18 with SinkhornSort · CIFAR-100 | Top-5 accuracy: 85.31 -> 86.94 | Improved |
| use only a top-5 loss instead of mixing top-1 and top-5 (ablation) | ResNet18 with SinkhornSort · CIFAR-100 | Top-1 accuracy: 61.27 -> 55.62 | Hurt |
| use only a top-5 loss instead of mixing top-1 and top-5 (ablation) | ResNet18 with SinkhornSort · CIFAR-100 | Top-5 accuracy: 85.31 -> 87.04 | Improved |
| replace softmax cross-entropy with a mixture of top-1 and top-5 loss | ResNeXt-101 32x48d WSL with SinkhornSort · ImageNet-1K | Top-5 accuracy: 97.795 -> 97.971 | Improved |
| replace softmax cross-entropy with a mixture of top-1 and top-5 loss | ResNeXt-101 32x48d WSL with DiffSortNets · ImageNet-21K-P | Top-1 accuracy: 39.29 -> 40.22 | Improved |
| increase top-5 weight from .5 to .75 (ablation) | Top-k DiffSortNets · ImageNet-21K-P | Top-1 accuracy: 40.22 -> 39.56 | Hurt |
| increase top-5 weight from .5 to .9 (ablation) | Top-k DiffSortNets · ImageNet-21K-P | Top-1 accuracy: 40.22 -> 38.48 | Hurt |
| use uniform weights for top-1 through top-5 rather than equal top-1/top-5 weights (ablation) | Top-k DiffSortNets · ImageNet-21K-P | Top-1 accuracy: 40.22 -> 39.69 | Hurt |
| use a top-20 component instead of a top-5 component (ablation) | Top-k DiffSortNets with equal top-1/top-k weights · ImageNet-21K-P | Top-1 accuracy: 40.22 -> 39.54 | Hurt |
| rank a much larger set of classes differentially, such as m=1,000 or more (ablation) | Differentiable sorting methods · ImageNet-1K | accuracy | Hurt |
Closest papers by meaning
- Analysis and Optimization of Loss Functions for Multiclass, Top-k, and Multilabel Classification (TPAMI 2017 · similarity 0.89)
- Loss Functions for Top-k Error: Analysis and Insights (CVPR 2016 · similarity 0.88)
- Smooth Loss Functions for Deep Top-k Classification (ICLR 2018 · similarity 0.88)
- ST$_k$: A Scalable Module for Solving Top-k Problems (NeurIPS 2024 · similarity 0.84)
- Cardinality-Aware Set Prediction and Top-$k$ Classification (NeurIPS 2024 · similarity 0.83)
- Differentiable Model Scaling using Differentiable Topk (ICML 2024 · 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.
