Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning Robust Representations by Projecting Superficial Statistics Out
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 |
|---|---|---|---|
| project out the texture-associated representation with HEX | image classifiers · MNIST-Rotation average | accuracy: 95.3 -> 95.8 | Helped |
| use HEX instead of the SURF-only baseline | Office features · Office D,W → A | accuracy: 0.112±0.008 -> 0.147±0.004 | Helped |
| use HEX instead of the SURF-only baseline | Office features · Office A,W → D and Office A,D → W | accuracy: 0.405±0.016 -> 0.343±0.030; 0.400±0.016 -> 0.378±0.034 | Hurt |
| replace NGLCM with a one-layer MLP (ablation) | HEX · facial-expression classification with nuisance background | testing accuracy under stronger train-test background correlation | Hurt |
| remove HEX projection and train with F_A (ablation) | HEX · facial-expression classification with nuisance background | testing accuracy under distribution shift | Hurt |
| add NGLCM texture features (ablation) | HEX · MNIST with independently attached patterns | testing accuracy | Mixed |
| add HEX projection (ablation) | pattern-robust MNIST classifier · MNIST with dependent pattern attachment, especially random-kernel test | testing accuracy | Helped |
| predict with projected logits F_L at test time (ablation) | HEX-ALL · Office features | accuracy: 0.343±0.030 -> 0.216±0.119; 0.147±0.004 -> 0.055±0.004; 0.378±0.034 -> 0.151±0.008 | Hurt |
| add the texture-only loss term with λ=1 (ablation) | HEX-ADV · synthetic datasets | testing performance | No clear effect |
| predict with F_L at test time (ablation) | HEX-ALL · synthetic datasets | testing performance | No clear effect |
| use NGLCM rather than MLP to extract texture-associated features (ablation) | texture and digit representations · mixed MNIST, SVHN, MNIST-M, and USPS | digit classification accuracy from extracted representations: 0.447±0.039 -> 0.161±0.008; 0.534±0.022 -> 0.142±0.023 | Helped |
| replace NGLCM with an MLP (ablation) | texture-feature extractor · mixed MNIST, SVHN, MNIST-M, and USPS | domain classification accuracy from extracted representations: 0.686±0.020 -> 0.738±0.018; 0.750±0.054 -> 0.687±0.029 | Mixed |
Closest papers by meaning
- Domain Invariant Representation Learning with Domain Density Transformations (NeurIPS 2021 · similarity 0.88)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.87)
- START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation (NeurIPS 2024 · similarity 0.86)
- Gradient-Guided Annealing for Domain Generalization (CVPR 2025 · similarity 0.86)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.85)
- Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize (ICML 2024 · similarity 0.85)
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.