multilayer perceptron

A multilayer perceptron, or MLP, is the most basic kind of neural network: rows of artificial neurons stacked one after another, with information flowing straight through from input to output. Picture an assembly line of decision-makers. The first row reads the raw input, each neuron makes its little weighted judgment and passes its number to the next row, which combines those judgments into more abstract ones, and so on until the final row delivers an answer. There are no loops and no shortcuts — just a clean, forward cascade, which is why it's also called a feedforward network.

What sets the MLP apart from its ancestor, the single perceptron, is the hidden layers in the middle — the rows that are neither input nor output. A lone perceptron can only carve data with one straight line and famously fails at simple patterns like XOR. Stack a hidden layer between input and output, add a nonlinear activation function at each neuron, and suddenly the network can bend and fold its decision boundaries into almost any shape. That leap, from one layer to many, is what made neural networks genuinely powerful.

Every neuron in one layer typically connects to every neuron in the next — this is the fully-connected layer, the MLP's signature. That density is its strength and its weakness: it can learn rich relationships among inputs, but it has no built-in sense of structure, so it treats a shuffled image the same as an ordered one. For data with spatial or sequential structure, specialized networks (convolutional or recurrent) usually do better. The MLP remains the conceptual foundation, though — understand it and you understand the skeleton beneath most of deep learning.