The Math Behind GANs
13 days ago
- #deep-learning
- #generative-models
- #machine-learning
- GANs involve a generator and discriminator competing to model data distributions.
- The discriminator's loss function aims to correctly classify real vs. fake data.
- The generator's loss function seeks to fool the discriminator by producing realistic data.
- Binary cross entropy is commonly used as the loss function for both models.
- Optimal discriminator performance is achieved when it balances detection of real and fake data.
- Training the generator minimizes the Jensen-Shannon divergence between real and generated data distributions.
- The original GAN formulation frames the training as a min-max optimization problem.
- Practical training alternates between updating the discriminator and generator parameters.
- Advanced GAN variants like Wasserstein GAN and CycleGAN build on these foundational concepts.