Advance AI MCE-E205

0 of 4 lessons complete (0%)
Exit Course

Deep Neural Network

Simulation AE and VAE

You don’t have access to this lesson

Please register or sign in to access the course content.

Autoencoder vs VAE Simulation

AE vs VAE Latent Space Interactive Simulation

Move your mouse over the “Latent Space” grids to see how the Decoder interprets your position.

Regular Autoencoder

Discrete points with empty space between them.

Latent Space (AE)
?

Decoder Output

Variational Autoencoder (VAE)

A continuous probability field. Everything is mapped.

Latent Space (VAE)
?

Decoder Output

What This Simulation Illustrates

1. In the Regular Autoencoder (AE) Panel

  • How it looks: The Latent Space is a scatter plot of discrete points (blue for “0,” red for “1”). Each input point maps to one fixed coordinate.
  • The Interactivity: Move your cursor over the clusters. The output easily decodes the digit.
  • The Defect (Gaps): Now, move your mouse into the empty white space between the clusters. The output box displays “⛶” (static). This is because the Regular AE has never seen this coordinate before; it is “off-manifold,” so its reconstruction is gibberish.

2. In the Variational Autoencoder (VAE) Panel

  • How it looks: The Latent Space is a smooth color gradient (blue on the left, blending into red on the right). Everything is filled in.
  • The Interactivity: As you move your mouse, the output doesn’t just jump between “0” and “1.” It transitions smoothly.
    • Far left (the “0” cloud) = “0”.
    • Far right (the “1” cloud) = “1”.
    • The Middle: As you slide your mouse from left to right, the “0” closes its shape and transforms smoothly into a “1” right before your eyes.
  • The Success: There are no gaps. Any random coordinate you click will result in a realistic-looking, synthetic “digit.” This is why VAEs are generative models.