[Paper Review] InfoGAN-CR: Disentangling Generative Adversarial Networks with Contrastive Regularizers

Zinan Lin, Kiran Koshy Thekumparampil+2 more|arXiv (Cornell University)|Jun 14, 2019

Generative Adversarial Networks and Image Synthesis17 citations

TL;DR

This paper introduces InfoGAN-CR, a self-supervised method for training disentangled generative adversarial networks using a contrastive regularizer inspired by latent traversal, achieving state-of-the-art disentanglement scores without ground-truth labels. It further proposes ModelCentrality, an unsupervised model selection scheme that identifies high-performing models via synthetic sample medoids, enabling effective hyperparameter tuning without access to labeled data.

ABSTRACT

Disentangled generative models map a latent code vector to a target space, while enforcing that a subset of the learned latent codes are interpretable and associated with distinct properties of the target distribution. Recent advances have been dominated by Variational AutoEncoder (VAE)-based methods, while training disentangled generative adversarial networks (GANs) remains challenging. In this work, we show that the dominant challenges facing disentangled GANs can be mitigated through the use of self-supervision. We make two main contributions: first, we design a novel approach for training disentangled GANs with self-supervision. We propose contrastive regularizer, which is inspired by a natural notion of disentanglement: latent traversal. This achieves higher disentanglement scores than state-of-the-art VAE- and GAN-based approaches. Second, we propose an unsupervised model selection scheme called ModelCentrality, which uses generated synthetic samples to compute the medoid (multi-dimensional generalization of median) of a collection of models. The current common practice of hyper-parameter tuning requires using ground-truths samples, each labelled with known perfect disentangled latent codes. As real datasets are not equipped with such labels, we propose an unsupervised model selection scheme and show that it finds a model close to the best one, for both VAEs and GANs. Combining contrastive regularization with ModelCentrality, we improve upon the state-of-the-art disentanglement scores significantly, without accessing the supervised data.

Motivation & Objective

• To address the challenge of training disentangled generative adversarial networks (GANs), which remain difficult despite advances in VAE-based methods.
• To develop a self-supervised contrastive regularizer that enforces disentanglement by leveraging the natural notion of latent traversal.
• To eliminate reliance on ground-truth labels for hyperparameter tuning by introducing an unsupervised model selection method, ModelCentrality.
• To achieve superior disentanglement performance on benchmark datasets without access to supervised annotations.

Proposed method

• Proposes a contrastive regularizer that encourages disentanglement by contrasting latent code perturbations during training, based on the principle of latent traversal.
• Uses a contrastive loss function that pulls together positive samples (small perturbations of the same latent code) and pushes apart negative samples (distant or unrelated codes).
• Applies the contrastive regularizer within a GAN framework to train a generator that produces disentangled samples while maintaining high-fidelity generation.
• Introduces ModelCentrality, an unsupervised model selection technique that computes the medoid (multi-dimensional median) of a set of generated synthetic samples across multiple models.
• Uses the medoid of synthetic samples to identify the model closest to the optimal disentangled representation, enabling hyperparameter selection without ground-truth labels.
• Combines the contrastive regularizer and ModelCentrality to train and select disentangled GANs end-to-end in a self-supervised manner.

Experimental results

Research questions

• RQ1 Can a contrastive regularizer based on latent traversal improve disentanglement in GANs without requiring ground-truth labels?
• RQ2 How effective is the proposed ModelCentrality method in selecting high-performing disentangled models compared to supervised baselines?
• RQ3 Can the combination of contrastive regularization and unsupervised model selection outperform state-of-the-art VAE- and GAN-based disentanglement methods?
• RQ4 To what extent can self-supervision mitigate the challenges of training disentangled GANs in the absence of labeled data?

Key findings

• InfoGAN-CR achieves state-of-the-art disentanglement scores on benchmark datasets, surpassing both SOTA VAE-based and GAN-based methods.
• The contrastive regularizer significantly improves disentanglement by encouraging meaningful, interpretable disentanglement through latent traversal.
• ModelCentrality successfully identifies models close to the optimal disentangled configuration, even without access to ground-truth labels.
• The combination of contrastive regularization and ModelCentrality enables effective hyperparameter tuning in a fully unsupervised setting.
• The method achieves superior disentanglement performance without relying on any supervised data, demonstrating the power of self-supervision in representation learning.