Architectures · beginner · concept 78 of 311 · Updated Nov 4, 2026
Multilayer Perceptron (MLP)
A multilayer perceptron is the simplest deep network: layers of units where each one connects to every unit in the next. It can approximate almost any function given enough units, yet ignores the structure of its input, which is why images and sequences moved to convolutions and attention. MLPs still sit inside nearly every modern model, including transformers.
Key terms
Fully connected layerHidden layerUniversal approximationWeightsForward pass
Learn these first
Where you meet it in the real world
Tabular prediction, the feed-forward blocks in transformers, classification heads
Guides and articles
Multilayer perceptrons ↗
Aston Zhang and others
Neural Networks and Deep Learning ↗
Michael Nielsen