Architectures · Established · Intermediate

Autoencoder

A network trained to reconstruct its own input through a narrow bottleneck, learning a compressed representation on the way.

What Autoencoder is

The encoder compresses, the decoder rebuilds. What survives the bottleneck is what the model considers essential, which makes autoencoders useful for compression, denoising and anomaly detection.

How it works

Training minimises reconstruction error. Variational autoencoders add a probabilistic latent space that can be sampled, and the latent space in modern image diffusion pipelines is produced by exactly this kind of encoder.

Why it matters

Latent-space diffusion models — the reason image generation is affordable — depend on a good autoencoder to shrink pixels into a compact space.

Common uses

→Anomaly detection
→Image denoising
→Latent spaces for diffusion models
→Dimensionality reduction

Strengths

✓Unsupervised
✓Flexible architecture

Watch for

✓Reconstruction quality may not equal usefulness
✓Latent space can be poorly structured

Continue exploring

Diffusion Model Dimensionality Reduction Embeddings Generative AI