Research map / Label-efficient and robust learning / Recognition and long-tail
Investigating the Histogram Loss in Regression
Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes, Esraa Elelimy, Martha White
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 l2 loss with HL-Gauss | MLP regressor · CT Position | Test MAE: 0.176 -> 0.098 | Improved |
| use Gaussian target distribution instead of Dirac delta (HL-OneBin) (ablation) | histogram loss · CT Position | Test MAE: 0.335 -> 0.098 | Improved |
| use Gaussian target instead of zero-bias projected target (ablation) | histogram loss · CT Position | Test MAE: 0.103 -> 0.098 | Improved |
| replace l2 with softmax output and per-bin l2 loss | MLP regressor · CT Position | Test MAE: 0.176 -> 0.105 | Improved |
| add target noise or gradient clipping to l2 | l2 regression · CT Position | Test MAE: 0.176 -> 0.152 | Improved |
| add uniform component to target distribution (HL-Uniform) (ablation) | histogram loss · CT Position | Test MAE | Hurt |
| use histogram loss with fixed random representation (ablation) | MLP last layer only · CT Position (random representation) | Test MAE: 9.305 -> 2.727 | Improved |
| initialize l2 network with HL-Gauss representation (ablation) | l2 regression · CT Position | Test MAE: 0.176 -> 0.185 | No clear effect |
| add distribution prediction as auxiliary task to l2 (Multi-Task) (ablation) | l2 regression · CT Position | Test MAE | No clear effect |
| add higher-moment prediction auxiliary tasks to l2 (ablation) | l2 regression · CT Position | Test MAE | No clear effect |
| anneal sigma from high to low during training (ablation) | HL-Gauss · Song Year and Bike Sharing | Test MAE | Improved |
| replace l2 with HL-Gauss for LSTM time-series prediction | LSTM · ETTh1 | Test MSE: 0.775 -> 0.6575 | Improved |
Closest papers by meaning
- Learning Deep Embeddings with Histogram Loss (NeurIPS 2016 · similarity 0.87)
- The Earth Mover’s Pinball Loss: Quantiles for Histogram-Valued Regression (ICML 2021 · similarity 0.87)
- Improving Deep Regression with Ordinal Entropy (ICLR 2023 · similarity 0.86)
- Generative Cross-Entropy: A Strictly Proper Loss for Data-Efficient Classification (NeurIPS 2026 · similarity 0.84)
- Engineering the Neural Collapse Geometry of Supervised-Contrastive Loss (Student Abstract) (AAAI 2024 · similarity 0.84)
- Dist Loss: Enhancing Regression in Few-Shot Region through Distribution Distance Constraint (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.
