Research map / Learning theory and algorithms / Matrix and tensor methods
DeepGUM: Learning Deep Robust Regression with a Gaussian-Uniform Mixture Model
Stephane Lathuiliere, Pablo Mesejo, Xavier Alameda-Pineda, Radu Horaud
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 the Gaussian-uniform mixture (DeepGUM) loss | VGG-16 regressor · FLD upper-body | average MAE (pixels): 15.80 -> 15.63 | Improved |
| replace Biweight loss with DeepGUM loss | VGG-16 regressor · FLD upper-body | average MAE (pixels): 17.88 -> 15.63 | Improved |
| replace Huber loss with DeepGUM loss | VGG-16 regressor · FLD upper-body | average MAE (pixels): 18.08 -> 15.63 | Improved |
| use Biweight instead of L2 loss | VGG-16 regressor · FLD upper-body | average MAE (pixels): 15.80 -> 17.88 | Hurt |
| add DeepGUM loss (instead of L2) | VGG-16 regressor · FLD lower-body | average MAE (pixels): 8.87 -> 8.53 | Improved |
| add DeepGUM loss (instead of L2) | VGG-16 regressor · FLD full-body | average MAE (pixels): 11.89 -> 11.73 | Improved |
| add DeepGUM loss (instead of L2) | VGG-16 regressor · McGill head pose (7-fold CV) | MAE (degrees): 8.60 -> 7.61 | Improved |
| add DeepGUM loss (instead of Biweight) | VGG-16 regressor · McGill head pose (7-fold CV) | MAE (degrees): 7.81 -> 7.61 | Improved |
| add DeepGUM loss (instead of L2) | VGG-16 regressor · CACD age estimation | MAE (years): 5.08 years of MAE (0.7 years better than L2) | Improved |
| add GUM mixture on top of Dex architecture (DexGUM) | Dex · CACD age estimation | MAE (years): p<0.001 | Improved |
| increase the share of uniformly generated outliers beyond 10% (l-UGO, g-UGO), DeepGUM vs L2/Biweight/Huber (ablation) | DeepGUM · facial landmark detection (LFW+NET, synthetic outliers) | failure rate: <10% comparable; larger amounts DeepGUM clearly superior | Improved |
| increase l-UGO outlier share to about 40% (ablation) | DeepGUM · facial landmark detection (l-UGO) | failure rate: breakdown point ~40% | Hurt |
Closest papers by meaning
- Nonlinear Regression via Deep Negative Correlation Learning (TPAMI 2019 · similarity 0.85)
- RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples (ICML 2024 · similarity 0.84)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.83)
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.83)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.83)
- Adaptive Robust Learning using Latent Bernoulli Variables (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.
