Definition
An autoencoder has an encoder that maps input to a lower-dimensional latent vector and a decoder that reconstructs the input. Training minimizes reconstruction error.
Variants include denoising, variational (VAE), and sparse autoencoders.
In simple terms
Summarize a book into bullet notes (encode), then rewrite the book from those notes (decode). Good notes capture what mattered.
Where you see it
- Anomaly detection when reconstruction fails on odd inputs.
- Dimensionality reduction and feature learning.
- VAEs for generative modeling.
How it works
1.Encode
Compress input to latent representation.
2.Decode
Reconstruct from the latent code.
3.Minimize error
Train so reconstruction matches the original.
Why it matters
- Autoencoders teach representation learning — a stepping stone to modern generative models.
Often confused
Autoencoders are just for compression like ZIP.
They learn semantic latents for ML tasks, not bit-perfect file compression.