Redirecting to original paper in 30 seconds...

Click below to go immediately or wait for automatic redirect

arxiv_ai 96% Match Research Paper ML Researchers,Computer Vision Engineers,Data Scientists 2 weeks ago

IB-GAN: Disentangled Representation Learning with Information Bottleneck Generative Adversarial Networks

generative-ai › gans
📄 Abstract

Abstract: We propose a new GAN-based unsupervised model for disentangled representation learning. The new model is discovered in an attempt to utilize the Information Bottleneck (IB) framework to the optimization of GAN, thereby named IB-GAN. The architecture of IB-GAN is partially similar to that of InfoGAN but has a critical difference; an intermediate layer of the generator is leveraged to constrain the mutual information between the input and the generated output. The intermediate stochastic layer can serve as a learnable latent distribution that is trained with the generator jointly in an end-to-end fashion. As a result, the generator of IB-GAN can harness the latent space in a disentangled and interpretable manner. With the experiments on dSprites and Color-dSprites dataset, we demonstrate that IB-GAN achieves competitive disentanglement scores to those of state-of-the-art \b{eta}-VAEs and outperforms InfoGAN. Moreover, the visual quality and the diversity of samples generated by IB-GAN are often better than those by \b{eta}-VAEs and Info-GAN in terms of FID score on CelebA and 3D Chairs dataset.
Authors (4)
Insu Jeon
Wonkwang Lee
Myeongjang Pyeon
Gunhee Kim
Submitted
October 23, 2025
arXiv Category
cs.CV
arXiv PDF

Key Contributions

Proposes IB-GAN, a novel GAN-based unsupervised model that integrates the Information Bottleneck framework to achieve disentangled representation learning. It constrains mutual information in an intermediate layer, leading to a more interpretable and disentangled latent space compared to InfoGAN.

Business Value

Enables the creation of more interpretable and controllable generative models, which can be valuable in applications like data augmentation, style transfer, and generating synthetic data with specific attributes.

Paper Metadata

Innovation Type

Novel Model Architecture

Deployment Feasibility

Moderate, requires expertise in GANs and representation learning for implementation and fine-tuning.

Limitations Addressed

Difficulty in achieving disentangled and interpretable latent spaces in GANs,Limitations of existing GAN optimization methods

Performance Gains

Achieves competitive disentanglement scores and outperforms InfoGAN.

Technical Tags

Disentangled Representation LearningGenerative Adversarial Networks (GANs)Information Bottleneck (IB)Unsupervised LearningLatent SpaceMutual InformationGenerator ArchitecturedSprites DatasetColor-dSprites DatasetInfoGANbeta-VAE

Research Topics

Representation LearningGenerative ModelsUnsupervised LearningDeep Learning ArchitecturesInformation Theory

Methods & Architectures

IB-GANInformation BottleneckGenerative Adversarial NetworksEnd-to-end trainingMutual Information Constraint IB-GANGANInfoGANbeta-VAE

Applications & Tasks

Computer Vision Data Representation Image Generation Learning disentangled representationsOptimizing GANs with Information BottleneckAchieving interpretable latent spaces Disentangled Representation LearningImage GenerationFeature Extraction

Datasets & Benchmarks

Datasets
dSprites, Color-dSprites
Benchmarks
Disentanglement scores (competitive with beta-VAEs, outperforms InfoGAN)
Disentanglement scores

Related Fields

Machine LearningDeep LearningInformation TheoryComputer Vision

Keywords

GANDisentangled RepresentationInformation BottleneckUnsupervised LearningLatent SpaceGenerative ModelsRepresentation LearningDeep LearningInfoGANbeta-VAEImage Generation

Academic Context

#Representation Learning#Generative Models#Unsupervised Learning#Deep Learning Architectures#Information Theory

Commercial Potential

Potential Products
Image Generation ToolsData Augmentation LibrariesFeature Extraction Modules
Target Industries
TechnologyGamingMediaE-commerce
Use Case Examples
Generating diverse synthetic images for training other modelsControlling specific attributes of generated images (e.g., shape, color)Learning meaningful representations from unlabeled image data
Competitive Edge
Offers an alternative approach to disentangled representation learning by integrating the Information Bottleneck principle into GANs, aiming for better interpretability and disentanglement than InfoGAN.
Market Opportunity
Growing market for generative AI and representation learning solutions.
Revenue Models
Licensing of the IB-GAN modelintegration into AI platforms.

Resource Requirements

Compute Needs
High, typical for training GANs, requiring significant GPU resources.
Data Requirements
Requires datasets suitable for representation learning, like dSprites and Color-dSprites.
Deployment Constraints
Training stability of GANs can be a challenge; inference speed depends on model complexity.
Scalability
Scalability depends on the specific GAN architecture and training infrastructure.

Production Readiness

Maturity Level

Research

Time to Market

2-4 years for robust productization.

View Full Paper Back to Papers