Autoencoders

Best for: Compression, denoising, anomaly detection

How it works

An autoencoder compresses each input $x$ into a low-dimensional code $z=E_\theta(x)$ through an encoder and reconstructs it as $\hat x=D_\phi(z)$, training the bottleneck to minimise reconstruction error $\|x-\hat x\|^2$. Forcing $z$ to be low-dimensional (or adding noise, as in a denoising autoencoder) makes the code capture the data’s dominant structure rather than copy the input. Variational autoencoders regularise $z$ with a KL term to a prior, $\mathcal{L}=\mathbb{E}_{q(z|x)}[\log p(x|z)]-D_{\mathrm{KL}}(q(z|x)\,\|\,p(z))$, enabling generative sampling.

Common fields

Cybersecurity · manufacturing · medical imaging