Overview

Generative Adversarial Nets (GANs):

• Motivation: Traditional generative models struggle with approximating many intractable probabilistic computations.

• Components:

  • Generative Model (G): Captures the data distribution.

  • Discriminative Model (D): Estimates the probability that a sample is from the training data rather than G.

    𝐺

• Advantages:

  • No need for Markov chains.
  • Only backpropagation is used for gradients.
  • No inference is required during learning.
  • Can incorporate a variety of factors and interactions.
  • Produces state-of-the-art log-likelihood estimates and realistic samples.

Conditional Generative Adversarial Nets (CGANs):

• Unconditioned Model Limitation: No control over modes of the data being generated.
• Conditioning: Directs data generation based on additional information y (e.g., class labels, parts of data for inpainting, data from different modalities).
• Demonstration: Constructing CGAN and empirical results on MNIST and MIR Flickr 25,000 datasets.

Related Work

Multi-modal Learning for Image Labeling:

• Supervised neural networks have been successful but struggle with extremely large numbers of predicted output categories and probabilistic one-to-many mappings.
• Leveraging additional information from other modalities (e.g., natural language corpora) can improve classification performance.
• Conditional probabilistic generative models can handle one-to-many mappings by using a conditional predictive distribution.

Conditional Adversarial Nets

Generative Adversarial Nets:

• Structure:
  • Generator (G): Maps prior noise distribution pz(z) to data space as G(z;θg).
  • Discriminator (D): Outputs the probability that x came from training data rather than G.
• Training: Both G and D are trained simultaneously in a two-player min-max game

Conditional Adversarial Nets (CGANs):