What Is an Autoencoder?
A network trained to compress its input down and then rebuild it — a simple task that forces it to learn what actually matters about the data.
Compress, then reconstruct
An autoencoder is trained to take an input, squeeze it through a narrow internal bottleneck far smaller than the original, and then reconstruct the original input as closely as possible from just that compressed version. Succeeding at this forces the compressed representation to capture the input's most essential structure, since there's no room to preserve irrelevant detail.
Why the bottleneck is the whole point
If the network could pass the input through unchanged, it would learn nothing useful — the narrow bottleneck is what forces genuine compression. What ends up in that compressed layer is, in effect, a learned, dense representation of the data, similar in spirit to an embedding.
What autoencoders are used for
Beyond compression, autoencoders are used for anomaly detection (data that reconstructs poorly is often unusual or faulty), noise removal, and as a building block inside larger generative systems, where a compressed latent representation is easier to work with than raw, high-dimensional data.
Frequently Asked Questions
What is an autoencoder?
A neural network trained to compress its input into a smaller internal representation and then reconstruct the original from that compressed form — forcing it to learn the input's essential structure.
What are autoencoders used for?
Data compression, anomaly detection (poor reconstructions often signal unusual data), noise removal, and as a component inside larger generative AI systems.