Radial Basis Function Network

A Radial Basis Function Network (RBFN) is a three-layer artificial neural network in which the hidden units apply radial basis functions, most commonly Gaussians, as their activation functions. In standard feedforward networks, activations depend on a weighted sum of inputs. In an RBFN, each hidden neuron instead computes a distance-based response: it measures how far an input vector lies from a learned center point and produces an output that decreases (or increases) monotonically with that distance. The final output layer then combines these responses through a linear weighted sum, making the overall mapping a linear combination of localized, radially symmetric basis functions.

Training an RBFN typically proceeds in two stages. First, the centers of the radial basis functions are determined, often through unsupervised methods like k-means clustering on the training data, or by selecting a subset of training points directly. Second, the output-layer weights are fitted using linear least squares, which is computationally cheap and avoids the vanishing gradient problems that can affect deep networks trained end-to-end. This decoupled training procedure gives RBFNs a speed advantage over multilayer perceptrons in many settings, and the linear output stage guarantees a unique, globally optimal solution for the weights given fixed centers.

RBFNs are well-suited to interpolation and function approximation tasks because each basis function acts as a local detector, responding strongly only to inputs near its center. This locality means the network can model nonlinear mappings while remaining interpretable, since the contribution of each hidden unit is spatially bounded and easy to visualize. Applications have included time-series forecasting, control systems, classification, and density estimation. They also connect to kernel methods and Gaussian processes, providing a bridge between neural network and statistical learning perspectives.

Deep learning has largely supplanted RBFNs for large-scale perception tasks, but they remain relevant where training data is limited, fast training is essential, or interpretability matters. Their theoretical properties, including universal approximation, convex output-layer optimization, and clear geometric interpretation, make them a useful reference point among neural network architectures.