Autoencoders

Autoencoder Architecture

Autoencoders learn compressed representations.

Architecture

Encoder: input → hidden (compressed). Decoder: hidden → output (reconstructed).

Loss: reconstruction error (input vs output). Bottleneck: smallest layer.

Types

Vanilla: fully connected. Convolutional: for images. Denoising: learns to denoise. Variational: generates new samples.

Applications

Dimensionality reduction. Anomaly detection (high reconstruction error = anomaly). Representation learning.

Key Takeaways

1. Autoencoders compress and reconstruct input
2. Bottleneck provides representation
3. Useful for anomaly detection