Multilayer perceptron
A multilayer perceptron (MLP) is a feedforward neural network with one or more hidden layers between input and output, typically built from fully connected layers. Nonlinear hidden activations let it represent relationships a single linear classifier cannot.
Also known as: MLP
A transaction classifier might receive a numeric vector containing amount, account age and purchase frequency. Each dense layer forms weighted sums and applies an activation function before passing values forward. The output can be a score for a class or a numerical prediction.
Training compares that output with a target through a loss function. Backpropagation computes gradients, and an optimizer updates weights. Nonlinear activations matter: stacking only linear transformations still produces one linear transformation overall.
Unlike a single threshold perceptron, an MLP can represent nonlinear boundaries in its input features. It does not include attention or recurrence by itself, though MLP blocks can be components of larger architectures such as transformers. This distinction helps when inspecting a model: a dense network over fixed features makes different assumptions from one designed to use image neighborhoods or sequence relationships. Choose and evaluate the architecture for the actual data.
Sources
- scikit-learn: Neural network models, supervised — Defines MLP hidden layers, nonlinear transformations, backpropagation and classification/regression uses.
Go deeper
- Andrej Karpathy: Building micrograd video
Build automatic differentiation and a small multilayer network from scratch.