Continuous Conditional Generative Adversarial Networks: Novel Empirical Losses and Label Input Mechanisms

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing conditional generative adversarial networks (cGANs) primarily handle categorical conditions, limiting their application to continuous regression labels.
• Conditioning on regression labels presents unique challenges, including data scarcity for specific labels and the inapplicability of conventional input mechanisms.
• These limitations hinder the effective generation of diverse, high-quality images based on continuous scalar conditions.

Purpose of the Study:

• To propose the first model, the continuous conditional generative adversarial network (CcGAN), for conditional generative modeling (CGM) with continuous regression labels.
• To address fundamental problems in applying cGANs to regression labels, specifically data scarcity and incompatible label input methods.
• To develop novel loss functions and label input mechanisms tailored for continuous conditional generation.

Main Methods:

• Reformulated existing empirical cGAN losses to be suitable for continuous regression scenarios, introducing hard vicinal discriminator loss (HVDL) and soft vicinal discriminator loss (SVDL).
• Proposed naive label input (NLI) and improved label input (ILI) mechanisms for incorporating regression labels into the generator and discriminator.
• Derived error bounds for discriminators trained with HVDL and SVDL, and introduced Sliding Fréchet Inception Distance (Sliding FID) as a novel evaluation metric.

Main Results:

• Developed four versions of CcGAN by combining proposed losses and label input mechanisms.
• Demonstrated CcGAN's capability to generate diverse, high-quality images conditional on regression labels across various benchmark datasets.
• CcGAN significantly outperformed traditional cGANs both visually and quantitatively in experiments.

Conclusions:

• The proposed CcGAN model is effective for conditional generative modeling with continuous regression labels.
• The novel loss functions and label input mechanisms successfully address the challenges associated with regression-based conditioning.
• CcGAN represents a significant advancement in generating conditional images, offering superior performance compared to existing cGAN approaches.