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

GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder

ICML 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot · 1 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 GenZSL, an inductive VAE that synthesizes unseen-class visual features by inducting from top-k similar seen classes using CLIP text embeddings, with class diversity promotion and boosting losses.

Problem

Generative zero-shot learning relies on expert-annotated class attributes and generates features from scratch, leading to poor sample quality and limited generalization to scenes; weak CLIP text class vectors make this worse.

Approach

Instead of imagining unseen-class features from Gaussian noise, an inductive VAE conditioned on CLIP class-name embeddings generates target-class features from referent seen-class samples (top-k similar classes), with class diversity promotion (removing the major SVD component of class vectors) and a target class-guided boosting loss.

Key results

• GenZSL reaches GZSL H of 87.4% on AWA2 versus 63.5% for f-VAEGAN and 57.4% H on CUB (best in red/blue table).
• CZSL accuracy of 73.5% on SUN and 92.2% on AWA2, the paper's reported best among compared generative methods.
• Reports about 60x faster training than f-VAEGAN on AWA2 and 10x on CUB.

Measured effects reported in the paper

Surprising findings

• Imagination-based generators (f-VAEGAN) degrade substantially when given weak CLIP text class vectors instead of expert attributes (SUN H 41.3% to 33.3%).

Limitations

• Relies on CLIP text embeddings and semantic similarity to select referent classes, so performance depends on the quality of those similarities.
• Several hyperparameters (lambda, top-k, N_syn) are tuned per dataset on the benchmarks.
• Results compare largely against imagination-based generative methods with varying backbones (some use CLIP features and others do not).

Experimental setup

Datasets: CUB SUN AWA2

Benchmarks: CUB SUN AWA2

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

Baselines: f-VAEGAN TF-VAEGAN CADA-VAE LisGAN HSVA CLSWGAN + DSP CLIP CoOp CoOp + SHIP ViFR VADS f-VAEGAN+DSP

Builds on: f-VAEGAN conditional VAE CLIP mixup

Closest papers by meaning

• Adaptive and Generative Zero-Shot Learning (ICLR 2021 · similarity 0.92)
• Data-Free Generalized Zero-Shot Learning (AAAI 2024 · similarity 0.91)
• ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · similarity 0.90)
• Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.90)
• Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 2024 · similarity 0.90)
• Visual-Augmented Dynamic Semantic Prototype for Generative Zero-Shot Learning (CVPR 2024 · similarity 0.89)