AI Glossary / Deep learning / Intermediate

Radial Basis Network

Definition

A radial basis network is a neural network that represents inputs through functions of their distance from selected centers. A common version uses Gaussian radial basis units in a hidden layer and combines their responses with weighted sums to produce predictions.

What it means

Each center is a reference point in the input feature space. A Gaussian unit responds most strongly near its center and less strongly farther away. Several such responses provide a set of features describing where an input lies relative to the centers.

The output layer combines these features to approximate a numerical function or support classification. The response of one unit is not automatically a class probability or a statement of confidence.

A traditional radial basis function, or RBF, network is usually shallow. Its nonlinear behavior comes from the basis functions, even when fitting the output weights is a linear problem. It is therefore related to neural learning without requiring a deep architecture.

How it works

For a Gaussian unit, a common response is exp(−||x − c||² / (2σ²)), where x is the input, c is the center and σ controls width. The distance depends on the feature representation and scaling.

Training can proceed in stages: choose centers, choose or estimate widths, then fit output weights. Centers might come from training examples, clustering or an incremental selection procedure. Some approaches optimize several components jointly.

With fixed centers and widths, the basis responses form a design matrix and the output weights can be fitted using linear methods, possibly with regularization. Width and center count affect the balance between local detail and smoother behavior. These choices should be assessed on validation data rather than selected only for minimum training error.

Three practical examples

A calibration model approximates a sensor's nonlinear response within an experimentally observed operating range.
A teaching classifier represents measurements by their distances from several prototype centers before combining the resulting features into class scores.
An engineering surrogate approximates a costly simulation using selected input-output examples, with separate checks near the boundaries of the sampled region.

Think of it this way

Think of overlapping pools of light placed at different positions. A point near a lamp receives a strong contribution from it; a distant lamp contributes less. Combining the pools produces a pattern across the space.

The analogy captures local Gaussian responses and overlap. Output weights can be negative, unlike ordinary light intensity, and a numerical feature space need not resemble physical geography. The centers are modeling choices, not necessarily meaningful real-world categories.

Worked example

Combining two local responses

Consider a one-dimensional teaching network with centers c1 = 0 and c2 = 2. Both widths are σ = 1. At input x = 1, the distance to each center is one, so both Gaussian responses are exp(−1/2), approximately 0.607.

Suppose the output has bias 0.5 and weights 2 and −1. Its prediction is 0.5 + 2 × 0.607 − 0.607, approximately 1.107.

The two positive basis responses have been combined using one positive and one negative weight. Neither response needed to be a probability. If x moves very far from both centers, both Gaussian responses approach zero and this particular network approaches its bias, 0.5.

That last behavior is a consequence of the stipulated architecture, not evidence that 0.5 is sensible outside the observed data range.

Illustrative case study

Fictional Elm Calibration builds an RBF model from temperature and vibration measurements. The first version uses the raw numbers directly. Because vibration is recorded on a much larger numerical scale, its differences dominate the distance calculation.

The team revisits units and feature scaling, fits the preprocessing using training records, and preserves the same transformation for later inputs. It then compares center counts and widths on separate validation observations.

A new operating condition lies far from all selected centers. Instead of accepting the network's output merely because it is a number, the team marks the condition as outside the evaluated range and collects additional reference measurements. The case illustrates how geometry and coverage influence an RBF model's usefulness.

Common misconceptions and mistakes

“Radial” means the function depends on distance from a center. It does not require inputs to be circular images or physical coordinates.

A Gaussian RBF network is not identical to every model using an RBF kernel. Kernel machines and RBF networks can share mathematical ingredients while differing in center selection, fitting and prediction structure.

Very narrow units can fit isolated examples while behaving poorly between them. Very broad units can struggle to represent local detail. Exact interpolation of training values is also not proof of accuracy on noisy or unseen observations. More centers can add cost and instability without improving useful prediction.

Business use

RBF networks can provide interpretable geometric building blocks for regression and approximation, particularly when meaningful distances can be defined. A project should document feature scales, center selection, width conventions, regularization and the domain in which performance has been tested.

Compare against simpler regression and other appropriate nonlinear methods using identical evaluation data. Include latency and memory as the number of centers grows. For a surrogate model, distinguish interpolation within supported regions from extrapolation into new operating conditions; a small training error says little about the latter.

Check your understanding

An RBF network fits every training point exactly. Can you conclude that its predictions between those points and beyond them are reliable?

Answer: No. Interpolation of observed values does not establish generalization. Width, center placement, noise, feature scaling and coverage all matter, and performance must be checked on relevant unseen inputs.

Related terms

Radial basis function
Regularization
Kernel method
Gaussian kernel
centre
bandwidth
function approximation
interpolation
feature scaling

Learn next

Learn distances, Gaussian curves and linear regression first. Then plot the responses of two or three basis functions and their weighted sum. Experiment with widths while holding the data fixed, and compare training fit with validation error and behavior outside the sampled interval.

Sources

Sources: Broomhead and Lowe, Multivariable Functional Interpolation and Adaptive Networks: https://sci2s.ugr.es/keel/pdf/algorithm/articulo/1988-Broomhead-CS.pdf Carnegie Mellon course notes, Radial Basis Function Networks: https://www.cs.cmu.edu/afs/cs/academic/class/46838-s99/classnotes/ch8-x4.pdf MathWorks, radial basis network design and spread behaviour: https://www.mathworks.com/help/deeplearning/ref/newrb.html MathWorks, Radial Basis Functions for Model Building: https://www.mathworks.com/help/mbc/model/radial-basis-functions-for-model-building.html

Previous: Radial Basis Function regression Next: Radon transform