Autoencoder
A neural network trained to reconstruct its input through a compressed bottleneck, learning efficient representations.
Definition
An autoencoder is a network that compresses its input to a low-dimensional code (the encoder) and then reconstructs the input from that code (the decoder). Training minimizes reconstruction error, forcing the code to retain the input's essential structure.
The size of the bottleneck controls what the model must keep: too wide and it copies the input without learning structure, too narrow and it loses essential detail. Variational autoencoders add a probabilistic structure that turns the code into a smooth, sampleable space.
The learned code is only as meaningful as the reconstruction task forces it to be; without constraints, a powerful autoencoder can learn a near-identity mapping that generalizes poorly. Bottlenecks, noise, and probabilistic structure are the tools that push the code to capture genuine structure. This makes autoencoders a flexible framework whose behavior depends heavily on the constraints the designer imposes.
Variants
- Denoising: reconstruct clean input from a corrupted version.
- Variational: learn a probabilistic latent space for generation.
- Sparse: encourage few active code units.
Why it matters
Autoencoders learn compact representations without labels, useful for dimensionality reduction, denoising, and anomaly detection, where inputs that reconstruct poorly are flagged as unusual. Variational autoencoders extend the idea to generative modeling.
Fusion connection
An autoencoder trained on normal simulated discharges flags anomalous plasma states by their high reconstruction error, a candidate early-warning signal for Hyperion studies.