Autoencoder
A neural network that compresses data into a compact representation, then reconstructs it.
An autoencoder is a type of neural network trained to copy its input to its output through a constrained internal representation called a latent space or bottleneck. The architecture consists of two components: an encoder that maps the input into a lower-dimensional latent representation, and a decoder that reconstructs the original input from that representation. By forcing information through this compressed bottleneck, the network learns to capture only the most essential structure in the data, discarding noise and redundancy. The training objective minimizes a reconstruction loss — typically mean squared error or cross-entropy — between the original input and its reconstruction.
The power of autoencoders lies in what the encoder learns to produce: a compact, structured representation of the data that can be used independently of the decoder. This makes autoencoders valuable for dimensionality reduction, anomaly detection, and feature extraction. Variants such as denoising autoencoders, which are trained to reconstruct clean inputs from corrupted versions, improve robustness and encourage the network to learn more meaningful representations. Sparse autoencoders add a regularization penalty to encourage only a small number of latent units to activate at once, promoting interpretable, disentangled features.
Variational autoencoders (VAEs), introduced in 2013, extended the framework by imposing a probabilistic structure on the latent space — typically a Gaussian distribution — enabling the model to generate new samples by sampling from that distribution. This made autoencoders a foundational tool in generative modeling, bridging unsupervised representation learning and data synthesis. More recently, autoencoder-style architectures have appeared in diffusion models and large-scale image generation systems, where they compress images into latent spaces for efficient processing.
Autoencoders matter because they offer a principled, unsupervised way to learn data representations without requiring labeled examples. They have been applied across domains including image compression, speech processing, drug discovery, and anomaly detection in industrial systems. Their conceptual simplicity — encode, compress, reconstruct — belies the richness of what they can learn, and they remain a core building block in modern deep learning pipelines.
