Research map / Learning theory and algorithms / Prediction and decision losses
Stochastic Loss Function
Qingliang Liu, Jinmei Lai
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 fixed cross-entropy with stochastic loss function (Gumbel-Softmax loss selection) | ResNet-20 · CIFAR-10 | accuracy: 0.9125 -> 0.9302 | Improved |
| replace fixed cross-entropy with SLF | DAC · ImageNet-10 clustering | ACC: 0.5272 -> 0.7345 | Improved |
| replace C-E loss with SLF | Transformer · IWSLT-14 De-En | BLEU: 34.01 -> 35.53 | Improved |
| learn YOLO loss weights with SLF | YOLO · PASCAL VOC 2007 | mAP: 63.4 -> 65.4 | Improved |
| replace C-E loss with SLF | ChebNets · DPP4 | squared correlation: 0.2650 -> 0.2793 | Improved |
| use Gumbel-Softmax selection instead of softmax over all losses (ablation) | SLF · CIFAR-10 clustering (DAC) | ACC: 0.5013 -> 0.5400 | Improved |
| increase sampling times K beyond 3-4 (ablation) | SLF · MNIST/CIFAR-10 | accuracy | Hurt |
| use SLF on imbalanced MNIST subsets (ablation) | DAC · imbalanced MNIST clustering | accuracy: more than 5% point | Improved |
| use learnable adaptive robust loss (ARLF) instead of handcrafted large-margin softmax | baseline loss · CIFAR-10 ResNet-8 | accuracy: 0.8866 -> 0.8646 | Hurt |
Closest papers by meaning
- Generalization Performance of Multi-pass Stochastic Gradient Descent with Convex Loss Functions (JMLR 2021 · similarity 0.85)
- Stochastic Gradient/Mirror Descent: Minimax Optimality and Implicit Regularization (ICLR 2019 · similarity 0.85)
- Asymmetric Loss Functions for Learning with Noisy Labels (ICML 2021 · similarity 0.85)
- How Does Loss Function Affect Generalization Performance of Deep Learning? Application to Human Age Estimation (ICML 2021 · similarity 0.85)
- Newton Losses: Using Curvature Information for Learning with Differentiable Algorithms (NeurIPS 2024 · similarity 0.85)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 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.
