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

ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning

ICLR 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot · 2 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 ZeroDiff, a diffusion-augmented generative zero-shot learning framework with supervised-contrastive instance-level representations and Wasserstein mutual-learning discriminators, aimed at robustness when seen-class training data is scarce.

Problem

Generative zero-shot learning methods degrade sharply when few training samples per seen class are available, attributed to spurious visual-semantic correlations learned by GAN discriminators.

Approach

Combine a diffusion forward process (to turn limited clean features into many noised samples), supervised-contrastive representations as instance-level semantics, and three discriminators (adversarial, diffusion, representation) trained with a Wasserstein-distance mutual-distillation loss weighted by the noise-to-data ratio.

Key results

• ZeroDiff reaches 87.3% ZSL T1 on AWA2, 87.5% on CUB and 77.3% on SUN, the best among compared embedding and generative methods reported.
• With 10% of seen-class training data on AWA2, ZeroDiff achieves 83.3% T1 and 77.0% H versus 73.1% T1 and 54.4% H for f-VAEGAN.
• On GZSL, ZeroDiff reports harmonic means of 81.4% (AWA2) and 81.6% (CUB), with the second-best at 79.3% and 81.1%.

Measured effects reported in the paper

Surprising findings

• Supervised-contrastive representations generally perform worse than cross-entropy features when used directly for generative ZSL, despite the paper using them as instance-level semantics.
• f-VAEGAN can underperform the vanilla f-CLSWGAN under limited data (e.g., SUN at 30% data: 72.0% vs 73.4% T1).

Limitations

• Relies on pre-defined attribute or text semantics and ResNet-101 features; comparisons with ViT-based methods use different backbones.
• Direct use of SC-based representations is worse than CE features, so the method depends on a separate representation generator and extra training stages.
• Diffusion uses only T=4 steps and the GAN-based training is complex with many discriminators and hyper-parameters.

Experimental setup

Datasets: AWA2 CUB SUN ImageNet-1K

Benchmarks: AWA2 CUB SUN

Metrics: top-1 accuracy unseen (T1) GZSL harmonic mean (H) seen accuracy (S) unseen accuracy (U)

Baselines: f-CLSWGAN f-VAEGAN TFVAEGAN SDVAE CEGAN DFCAFlow DSP DML VADS ICIS ReZSL PSVMA

Builds on: f-VAEGAN DDGAN supervised contrastive loss VADS PaCo

Closest papers by meaning

• ZeroMamba: Exploring Visual State Space Model for Zero-Shot Learning (AAAI 2025 · similarity 0.92)
• Progressive Semantic-Guided Vision Transformer for Zero-Shot Learning (CVPR 2024 · similarity 0.91)
• Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning (CVPR 2024 · similarity 0.91)
• GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder (ICML 2025 · similarity 0.90)
• Epsilon: Exploring Comprehensive Visual-Semantic Projection for Multi-Label Zero-Shot Learning (AAAI 2025 · similarity 0.90)
• Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.90)