Why this phase exists

Every model so far has learned to predict something — a class, a number, the next word in a sentence. This phase flips the question around: can a network learn to generate brand-new data that looks like it belongs to the training set?

Two very different families answer that question:

• Autoencoders learn to copy their input to their output through a narrow bottleneck. That constraint forces them to discover a compact, meaningful representation of the data — the latent code (or coding).
• GANs (Generative Adversarial Networks) pit two networks against each other: a generator that fakes data, and a discriminator that tries to catch the fakes. Neither one ever “solves” the problem alone — they get better by competing.

In this phase, you’ll learn:

• how undercomplete, stacked, denoising, sparse, and convolutional autoencoders work, and how they’re used for dimensionality reduction and unsupervised pretraining
• how a variational autoencoder (VAE) turns the latent space into a smooth, samplable probability distribution so you can generate brand-new data
• how a GAN’s generator and discriminator are trained in two competing phases, why that training is famously unstable, and how DCGANs stabilize it with convolutions

Phase 6 flow

  flowchart TD
  A["Input data"] --> B["Autoencoder: encoder to bottleneck to decoder"]
  B --> C["Reconstruction"]
  B --> D["Variational Autoencoder: probabilistic latent space"]
  D --> E["Sample z, then decode: new data"]
  A --> F["GAN: Generator vs Discriminator"]
  F --> G["Adversarial training loop"]
  G --> H["Realistic generated data"]

Suggested practice dataset

Fashion MNIST is the perfect sandbox for this phase — it’s exactly what Géron uses in the book. It’s small and grayscale, so an autoencoder or a small GAN trains in minutes on a laptop CPU, yet it’s rich enough that you can actually see meaningful reconstructions and generations.

Next

Start with Autoencoders — the encoder/decoder bottleneck that everything else in this phase builds on.