Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Generalized Zero-Shot Classification via Semantics-Free Inter-Class Feature Generation
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 semantics-free generalized zero-shot classification method for ordered clinical classes that generates unseen-class features by mixing neighboring seen-class features under a learned ordinal Gaussian constraint.
Problem
Generalized zero-shot learning for medical image classification (glioma subtyping, AD, diabetic retinopathy) where semantic attributes are hard to define or align with visual features.
Approach
Exploits the inherent risk ordering of clinical classes: an inter-class feature augmentation (IFA) module imposes a pre-defined joint conditional Gaussian ordering on learned features via a KL ordinal loss, then synthesizes unseen-class features by averaging features of adjacent seen classes, turning GZSL into supervised learning without any semantic information.
Key results
- On the in-house glioma task, IFA reaches HM 71.6 and AUC 0.928 versus 66.8 HM and 0.729 AUC for the best semantic-based baseline SD-GZSL.
- On ADNI, IFA achieves HM 82.4 and AUC 0.973, outperforming FREE (HM 76.0, AUC 0.972) with Delong-test p<0.05 across datasets.
- Ablation shows removing the distinctiveness loss drops glioma HM from 71.6 to 64.4 and APTOS-1 HM from 65.3 to 40.5-level (comparison reported in Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IFA instead of the strongest reported GZSL baseline | GZSL classification · Glioma | HM: 66.8 -> 71.6 | Helped |
| use IFA instead of the strongest reported GZSL baseline | GZSL classification · ADNI | HM: 76.0 -> 82.4 | Helped |
| use IFA instead of the strongest reported GZSL baseline | GZSL classification · APTOS-1 | HM: 63.5 -> 66.5 | Helped |
| use IFA instead of the strongest reported GZSL baseline | GZSL classification · APTOS-2 | HM: 54.4 -> 56.6 | Helped |
| use IFA instead of the strongest reported GZSL baseline | GZSL classification · APTOS-3 | HM: 57.5 -> 70.4 | Helped |
| add distinctiveness loss LDIS (ablation) | IFA · Glioma | HM: 64.4 -> 71.6 | Helped |
| add distinctiveness loss LDIS (ablation) | IFA · ADNI | HM: 78.4 -> 82.4 | Helped |
| add distinctiveness loss LDIS (ablation) | IFA · APTOS-1 | HM: 40.5 -> 65.3 | Helped |
| add distinctiveness loss LDIS (ablation) | IFA · APTOS-2 | HM: 43.3 -> 56.6 | Helped |
| add distinctiveness loss LDIS (ablation) | IFA · APTOS-3 | HM: 46.1 -> 70.4 | Helped |
| add the IFA module with ordinal constraint rather than train without IFA (ablation) | IFA · Glioma, ADNI, and APTOS | ordinal feature distribution | Helped |
Surprising findings
- Ablation without the distinctiveness loss sometimes yields higher seen accuracy on APTOS-1 and ADNI, showing a trade-off with unseen accuracy.
Limitations
- Only works for unseen classes that have neighboring seen classes; edge classes at the ends of the ordering are poorly constrained.
- Relies on the assumption that clinical classes are inherently ordered by risk.
- Evaluated only on a small in-house glioma dataset with few astrocytoma samples (18 patients).
Experimental setup
Datasets: in-house glioma MRI dataset (726 patients) ADNI APTOS 2019
Benchmarks: in-house glioma dataset ADNI APTOS
Metrics: seen accuracy unseen accuracy harmonic mean AUC
Baselines: DFZSL HSVA CE-GZSL FREE SD-GZSL fully supervised
Builds on: DFZSL HSVA CE-GZSL FREE SD-GZSL
Closest papers by meaning
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.88)
- Data-Free Generalized Zero-Shot Learning (AAAI 2024 · similarity 0.87)
- Class Semantic Attribute Perception Guided Zero-Shot Learning (AAAI 2025 · similarity 0.86)
- Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.85)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.85)
- ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · 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.