Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition

AAAI 2024 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot · 26 citations (OpenAlex)

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 D3GZSL, an end-to-end framework that plugs into generative GZSL models (f-CLSWGAN, CE-GZSL, TF-VAEGAN, DDGAN) and uses in-distribution dual-space distillation and out-of-distribution batch distillation to reduce seen-class bias.

Problem

Generalized zero-shot learning models trained with generated unseen features are biased toward seen classes because the generator mainly models the seen-data distribution; existing OOD-based fixes are two-stage and accumulate errors.

Approach

Treat real seen features as in-distribution and generated unseen features as out-of-distribution, then train a teacher-student pair end-to-end with batch-wise embedding and logit distillation (ID2SD) plus a low-dimensional OOD confidence representation with batch-wise similarity distillation (O2DBD), yielding one unified classifier instead of two expert classifiers plus an OOD detector.

Key results

• Applied to f-CLSWGAN on CUB, H rises from 49.7 to 56.5 and on FLO from 65.6 to 71.7 (per-dataset gains reported as 6.8 and 6.1 points).
• CE-GZSL+D3GZSL reaches H of 70.5 on AWA1 and 67.8 on CUB, the best H among compared methods on those datasets.
• Ablation on CUB with f-CLSWGAN: baseline H 54.7, ID2SD alone 55.6, O2DBD alone 55.1, both 56.5.

Measured effects reported in the paper

Surprising findings

• The idealized two-stage OOD pipeline (IV-TS) with perfect OOD separation outperforms the proposed method on several datasets, while the practical two-stage method performs worse due to error accumulation.

Limitations

• Gains on AWA2 are small or negative for some baselines (e.g., TF-VAEGAN on AWA2 H 66.6 to 66.8, DDGAN on FLO U tradeoffs), and mixed on seen accuracy.
• Hyperparameters such as batch size and generated-sample counts differ per dataset and were set empirically.

Experimental setup

Datasets: AWA1 AWA2 CUB FLO

Benchmarks: AWA1 AWA2 CUB FLO

Metrics: top-1 accuracy on seen classes (S) top-1 accuracy on unseen classes (U) harmonic mean (H)

Baselines: cycle-CLSWGAN CADA-VAE LisGAN IZF SE-GZSL TDCSS DUET GKU f-CLSWGAN TF-VAEGAN CE-GZSL DDGAN

Builds on: f-CLSWGAN CE-GZSL TF-VAEGAN Maximum Softmax Probability Energy-based OOD

Closest papers by meaning

• Meta-Learning for Generalized Zero-Shot Learning (AAAI 2020 · similarity 0.91)
• Contrastive Embedding for Generalized Zero-Shot Learning (CVPR 2021 · similarity 0.91)
• Data-Free Generalized Zero-Shot Learning (AAAI 2024 · similarity 0.91)
• Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.91)
• GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder (ICML 2025 · similarity 0.90)
• ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · similarity 0.90)