Back to Projects
Generative Models

MNIST GAN

Generative Adversarial Network (GAN) for generating handwritten numbers from random noise.

Course: Neural NetworksCo-authors: Diego Quezada
MNIST GAN

Objectives

  • 1Implement a GAN to generate handwritten numbers from initial random noise.
  • 2Compare results of random noise generated by two numbers and the average of both.

Conclusions

  • The GAN generates recognizable digits after 50 epochs, with discriminator loss stabilizing around 0.5.
  • Training stability requires balancing discriminator and generator learning rates to prevent mode collapse.
  • Interpolating between two latent vectors produces smooth transitions with hybrid digit characteristics.

Technologies

  • Keras
  • NumPy
  • Matplotlib