Is an MLP always deep?

Terminology varies. An MLP can have a single hidden layer; deep usually indicates multiple learned layers or substantial depth.

Can an MLP handle images?

Yes, if represented appropriately, but a plain dense network does not automatically exploit image structure as specialized architectures do.

Does it remember previous inputs?

A basic feedforward MLP has no recurrent state. History must be represented in its inputs or handled by other system components.

AI Glossary / Deep learning / Intermediate

Multilayer perceptron

Definition

A multilayer perceptron, or MLP, is a feedforward neural network built from layers of connected computational units. Hidden layers usually combine weighted inputs with nonlinear activation functions, enabling the network to represent relationships beyond a single linear rule.

What it means

An MLP transforms an input vector into an output through a sequence of learned calculations. The input might contain product measurements, encoded customer activity or features produced by another model. The output might be a predicted quantity or class scores.

In ordinary usage, an MLP has an input stage, at least one hidden layer and an output layer, with dense connections between successive layers. Layer-counting conventions vary: some descriptions count only layers with learned parameters, while others include the input layer.

The name comes from the history of neural networks, but a modern MLP is not simply a collection of the original hard-threshold perceptrons. Its units often use differentiable or piecewise-differentiable functions suited to gradient-based training.

How it works

For each hidden unit, the network multiplies incoming values by weights, adds a bias and applies an activation function. The next layer receives those transformed values. Nonlinear activations matter: a chain of affine transformations without intervening nonlinearity can be collapsed into one affine transformation.

The final layer matches the task. A regression model might return a continuous value, while a classifier might convert output scores into class probability estimates. Such estimates still require evaluation and may need calibration.

Training compares predictions with targets using a loss function. Backpropagation calculates gradients; an optimizer uses them to update parameters. These are related but distinct operations.

Architecture, feature scaling, learning settings and regularization influence the result. More units can increase representational capacity and also make overfitting easier. Validation data help choose settings without treating training performance as evidence of future accuracy.

Three practical examples

A maintenance team predicts a continuous vibration measure from operating conditions using an MLP regressor.
A document workflow classifies records from fixed-length features extracted by an earlier processing stage.
A larger neural system uses an MLP component to transform a learned representation into task-specific output scores.

Think of it this way

Think of several adjustable mixing panels connected in sequence. Each panel combines incoming signals and applies a transformation before passing its outputs onward. Adjusting the controls changes the overall input-to-output behavior.

The analogy explains layered computation, but these panels do not understand the business meaning of their signals. Training adjusts numerical parameters to reduce a defined loss; it does not supply common sense or guarantee that the learned relationship remains valid outside the training setting.

Worked example

Following a small network forward

Consider two inputs, x1 = 2 and x2 = 1, and two hidden units with ReLU activation, which returns the larger of zero and its argument. Choose illustrative hidden calculations h1 = ReLU(x1 − x2) and h2 = ReLU(−x1 + 2x2).

The hidden values are h1 = 1 and h2 = 0. Let a linear output be y = 3h1 + h2 + 0.5. The prediction is therefore 3.5.

For inputs x1 = 1 and x2 = 2, the hidden values become 0 and 3, giving the same output of 3.5 through a different internal pattern. These hand-selected weights illustrate inference, not a completed learning experiment. Training would need examples, a loss and parameter updates. The example also shows why equal predictions do not imply identical internal representations.

Illustrative case study

A fictional service company, Birch Repairs, trains an MLP to estimate job duration from equipment and booking features. Its training errors look small, but performance on later bookings is disappointing.

The team discovers that some records contain a completion code entered only after the repair. Removing that unavailable feature changes the problem substantially. It then fits preprocessing using training data, compares the network with a simple baseline and evaluates later jobs without using their outcomes during tuning.

The project becomes a test of whether the available booking information supports useful predictions. The company does not assume that adding hidden layers can repair a misleading target, leaked information or a changed repair process.

Common misconceptions and mistakes

Every additional layer makes a model better. Extra capacity can increase data, tuning and computation requirements without improving the relevant result.
Backpropagation is the optimizer. It computes gradients that an optimizer can use.
Dense connections mean the model knows every relationship. They permit numerical interactions; useful relationships still have to be learned.
A probability-like output is necessarily well calibrated. Its numerical form alone does not establish reliability.

Business use

MLPs are useful general-purpose function approximators and common components inside larger AI systems. They can support structured-data prediction when their performance and operating costs justify their use.

A business review should compare the complete pipeline with suitable alternatives, including simpler models. Ask what inputs exist at prediction time, how missing values are handled, whether performance is stable across important groups and how often retraining is needed. Record the model and preprocessing together: applying different scaling in production can change the meaning of the network's inputs even when its weights remain unchanged.

Check your understanding

Answer: A network contains five affine layers but no nonlinear activations between them.

Does its depth alone give it a nonlinear input-to-output mapping?

Answer: No. Those affine transformations can be combined into one affine transformation. Nonlinear operations are needed for that particular source of additional expressiveness.

Related terms

Learn next

Calculate the worked network for another input pair and identify which ReLU units become zero. Then compare a linear model with a small MLP on the same held-out task, keeping preprocessing and evaluation rules consistent.

Linear model

Sources

Goodfellow, Bengio and Courville, Deep Learning, chapter 6, Deep Feedforward Networks: https://www.deeplearningbook.org/contents/mlp.html Scikit-learn, supervised neural-network models and practical guidance: https://scikit-learn.org/stable/modules/neural_networks_supervised.html

Previous: Multilabel indicator matrix Next: Multilevel Monte Carlo method