What is an autoencoder?

An autoencoder is a neural network trained to reconstruct its input after first compressing it into a smaller internal representation.

It has two main parts:

Input
  ↓
Encoder
  ↓
Latent representation
  ↓
Decoder
  ↓
Reconstructed input

How it works

The encoder transforms the input into a compact representation, often called a latent vector or bottleneck.

The decoder tries to rebuild the original input from that representation.

Training minimizes the difference between:

Original input
and
Reconstructed input

Because the model must reconstruct the input through a constrained representation, it can learn useful structure in the data.

Why are autoencoders useful?

Autoencoders are commonly used for:

• Dimensionality reduction
• Representation learning
• Denoising
• Anomaly detection
• Data compression research

For example, a denoising autoencoder can receive a noisy image and learn to reconstruct a cleaner version.

Important point

An autoencoder is not simply copying the input. Its usefulness comes from the constraints placed on the internal representation, such as a small bottleneck, noise, sparsity, or other regularization.

Different variants, such as variational autoencoders (VAEs), change how the latent representation is learned and can also be used for generative modeling.

In short, an autoencoder learns to encode data into a useful internal representation and decode that representation back into a reconstruction of the original input.