Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Class Distribution Shifts in Zero-Shot Learning: Learning Robust Representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a parametric model showing ERM zero-shot representations overfit to an unknown attribute's training mixture, and an algorithm that builds synthetic environments via hierarchical class subsampling and penalizes performance (AUC) variance across them.
Problem
Zero-shot verification/recognition models trained with metric losses become non-robust when the class distribution shifts through an unknown attribute (e.g., age, gender, hair color) between training and test classes.
Approach
Create artificial environments by sampling small subsets of training classes (so attribute mixtures vary), and penalize the variance of per-environment soft-AUC (VarAUC), applying OOD environment-balancing ideas to zero-shot metric learning.
Key results
- On CelebA and ETHEC, the VarAUC penalty improves over ERM under class distribution shift with one-sided paired t-test p-values below 0.04 (FDR-adjusted).
- In simulations, the proposed method improves AUC under distribution shift without degrading in-distribution performance compared with ERM.
- VarREx reaches high AUC faster than VarAUC, but VarAUC attains higher overall AUC; IRM converges noisily.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hierarchical sampling and VarAUC environment balancing | Zero-shot representation learning · Simulation, shifted distribution | AUC | Helped |
| Add hierarchical sampling and VarAUC environment balancing | Zero-shot representation learning · Simulation, in-distribution | AUC | No clear effect |
| Add hierarchical sampling and VarAUC environment balancing | Zero-shot representation learning · ETHEC, shifted species distribution | AUC | Helped |
| Add hierarchical sampling and VarAUC environment balancing | Zero-shot representation learning · CelebA, shifted blond-hair distribution | AUC | Helped |
| Use VarAUC rather than IRM, CLOvE, or VarREx for environment balancing (ablation) | Hierarchical-sampling zero-shot representation learning · Simulation and real-world datasets with distribution shifts | AUC | Helped |
| Replace VarAUC with VarREx (ablation) | Hierarchical-sampling zero-shot representation learning · Simulation, shifted distribution | AUC | Hurt |
| Replace VarAUC with IRM (ablation) | Hierarchical-sampling zero-shot representation learning · Simulation | AUC | Mixed |
| Remove the environment-balancing penalty while retaining hierarchical sampling (ablation) | Hierarchical-sampling zero-shot representation learning · Distribution-shift scenario | AUC | Hurt |
| Add dropout regularization (ablation) | Zero-shot representation learning · Distribution-shift scenario | AUC | No clear effect |
| Add L2-norm regularization (ablation) | Zero-shot representation learning · Distribution-shift scenario | AUC | No clear effect |
Surprising findings
- Standard regularization such as dropout and L2 norm did not yield notable improvements under distribution shift.
- Theory shows that the optimal ERM representation under the shift can prioritize features that separate only the majority training type.
Limitations
- Experiments only address shifts in a single binary attribute.
- Shifts with attributes correlated with other covariates are not studied.
- Parametric model is linear Gaussian, with analysis restricted to diagonal weights.
Experimental setup
Datasets: ETHEC CelebA synthetic Gaussian simulations
Benchmarks: ETHEC CelebA simulated parametric model
Metrics: AUC percentage change vs ERM
Baselines: ERM hierarchical sampling baseline IRM penalty CLOvE penalty VarREx penalty
Builds on: invariant risk minimization VarREx CLOvE contrastive loss
Closest papers by meaning
- Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.88)
- Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.88)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.87)
- Zero-Shot Robustification of Zero-Shot Models (ICLR 2024 · similarity 0.87)
- OTTER: Effortless Label Distribution Adaptation of Zero-shot Models (NeurIPS 2024 · similarity 0.87)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.87)
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.