I Use This When...
<!-- Practical use case -->History
Hinton & Salakhutdinov (1986/2006). The idea: compress data through a bottleneck, then reconstruct it. What survives the bottleneck is the essential information.
Why It Exists
PCA can only find linear relationships. An autoencoder uses neural networks for non-linear compression — it can learn much richer representations.
How It Works
Visual Intuition
<!-- 3B1B-style animation description -->Step by Step
<!-- Algorithm walkthrough -->Code
# Implementation sketch
The Math Inside
Encoder: z = f(x), Decoder: x' = g(z). Loss: L = ||x - x'||^2. Bottleneck z has fewer dimensions than x. Variational Autoencoder adds: z ~ N(μ, σ).
Math Prerequisites
<!-- Links to math wiki -->Related
- PCA — Linear version
- MLP & Backprop — How it trains
- GAN — Alternative generative approach
- Diffusion — Modern generative approach