Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · Multi-layer perceptron

The Multi-Layer Perceptron: Architecture, Backpropagation and Research Applications

The multi-layer perceptron is the architecture through which the training of neural networks became a solved engineering problem rather than a theoretical impasse. The algorithm that made this possible is backpropagation.

Pubrica AcademyWritten by a data-analytics reviewer Reviewed bya biostatistician Published10 October 2026 Last checked10 October 2026 Reading time9 minutes Sources1 peer-reviewed

A multi-layer perceptron (MLP) is a feedforward artificial neural network consisting of at least three layers of nodes (an input layer, one or more hidden layers and an output layer), in which every node in one layer is connected to every node in the next; the nonlinear activation functions of the hidden layers enable the network to represent relationships that a single-layer model cannot.

Learning objectives

  • The limitations of the single-layer perceptron and the capabilities gained by adding hidden layers.
  • The nature of backpropagation and the reasons it made the training of multi-layer networks practical.
  • The role of the activation function in enabling a multi-layer perceptron (MLP) to represent nonlinear relationships.
  • The research tasks for which an MLP is, and is not, an appropriate tool.

Common misconceptions

  • “A perceptron and a multi-layer perceptron are the same thing with a bigger name.” A single perceptron can separate only linearly separable data; it is the addition of hidden layers with nonlinear activation functions that enables a multi-layer perceptron to represent nonlinear relationships.
  • “Neural networks are a recent invention.” The backpropagation algorithm that made multi-layer perceptrons trainable in practice was described by Rumelhart, Hinton and Williams in 1986, and the underlying concept of networks of threshold units is older still.
  • “More hidden layers always means a better model.” Greater depth increases representational capacity but also increases the amount of data and regularisation required to avoid overfitting; no general rule establishes that additional layers improve performance on a given dataset.

Limitations of the single-layer perceptron

A single-layer perceptron computes a weighted sum of its inputs and passes it through a threshold, producing a decision boundary that is a straight line (or a hyperplane in higher dimensions). Consequently, a single perceptron can correctly classify only linearly separable data and cannot represent relationships such as XOR, in which no single straight line separates the classes. A multi-layer perceptron overcomes this limitation by introducing one or more hidden layers between input and output: each hidden node computes its own weighted sum and nonlinear activation, and the combination of many such nodes across layers enables the network to approximate considerably more complex, nonlinear decision boundaries.

Backpropagation and the practical training of multi-layer networks

The recognition that multi-layer networks could, in principle, represent nonlinear functions did not immediately make them trainable in practice, because no efficient method existed for adjusting the weights of hidden layers on the basis of the network’s output error. Rumelhart, Hinton and Williams, in their 1986 Nature paper, described the backpropagation algorithm: a method for efficiently computing how a small change in any weight in the network, including weights in hidden layers, affects the overall output error, by propagating the error backward from the output layer through each preceding layer using the chain rule of calculus1. This provided a practical and efficient means of training networks with hidden layers by gradient descent, and it is the specific contribution that made multi-layer perceptrons a usable tool rather than a theoretical possibility.

Role of the activation function

The nonlinear activation function applied at each hidden node (historically a sigmoid function; in modern practice often a rectified linear unit) is what gives the network its capacity to represent nonlinear relationships; a network composed entirely of linear operations, irrespective of its depth, still computes an overall linear function of its inputs. The description of backpropagation by Rumelhart, Hinton and Williams requires this nonlinearity to be differentiable, which is why the choice of activation function is a structural requirement of the training procedure rather than a minor detail1.

Appropriate and inappropriate research applications of the multi-layer perceptron

An MLP is a general-purpose, supervised function approximator: given labelled examples of inputs and desired outputs, it learns weights that map one to the other, which makes it suitable for classification or regression tasks on tabular, fixed-size feature inputs in which the relationship between features and outcome is expected to be nonlinear. It is not inherently suited to inputs with sequential or spatial structure (free text, images, time series) in the way that architectures designed specifically for those data types are, and, as with any supervised model, it requires a labelled dataset that is sufficiently large relative to its number of parameters to be trained without overfitting.

Key points

Key points

A multi-layer perceptron is a feedforward neural network whose hidden layers and nonlinear activation functions let it represent relationships a single-layer perceptron cannot, and it only became practically trainable once backpropagation (Rumelhart, Hinton and Williams, 1986) provided an efficient way to adjust hidden-layer weights from the output error.

Practical application

Practical application

  1. The need for a nonlinear model should be confirmed for the task. Before an MLP is adopted, authors should assess whether a simpler and more interpretable model (such as logistic regression) already fits the relationship adequately; additional complexity should be justified by a measured gain rather than assumed.
  2. Network depth and size should be matched to the size of the dataset. A larger or deeper MLP has more parameters to fit; without sufficient labelled examples relative to the number of parameters, overfitting rather than improved performance is to be expected.
  3. The architecture and training details required for a reviewer to assess the model should be reported. Authors should state the number of hidden layers and nodes, the activation function, the training algorithm and regularisation used, and the validation approach, not only the final accuracy figure.

Frequently asked questions

A multi-layer perceptron with many hidden layers is one example of a deep learning architecture, but the term deep learning generally refers to a broader family of architectures, including those (such as convolutional or recurrent networks) designed specifically for structured inputs such as images or sequences, for which a basic MLP is not designed.

Their Nature paper is the widely cited source that popularised an efficient, general method for training multi-layer networks by propagating error gradients backward through hidden layers, and it is therefore the standard citation for the technique in this context.

Yes. The same architecture, with an appropriate output layer and loss function, can be trained to predict a continuous outcome (regression) or a class label (classification).

References

  1. Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature. 1986;323:533-536. read 10 October 2026. View on Nature

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). The Multi-Layer Perceptron: Architecture, Backpropagation and Research Applications. Pubrica. https://pubrica.com/academy/concepts-definitions/multi-layer-perceptron/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

MLP reporting checklist for a manuscript

For use when describing a multi-layer perceptron model in a methods section.

[ ] Number of hidden layers and nodes per layer stated
[ ] Activation function(s) stated
[ ] Training algorithm and loss function stated
[ ] Regularisation / dropout approach stated
[ ] Train/validation/test split or cross-validation scheme stated
[ ] Dataset size relative to model parameter count discussed

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s data-analytics team builds, validates and reports multi-layer perceptron and other machine-learning models for research manuscripts, with architecture and training details described to the standard expected by a statistical reviewer.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Methodological support for reporting neural network models in research

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.