Multi-layer networks of threshold logic units offer an attractive framework for the design of pattern classification systems. A new constructive neural network learning algorithm (DistAl) based on inter-pattern distance is introduced. DistAl uses spherical threshold neurons in a hidden layer to find a cluster of patterns to be covered (or classified) by each hidden neuron. It does not depend on an iterative, expensive and time-consuming perceptron training algorithm to find the weight settings for the neurons in the network, and thus extremely fast even for large data sets. The experimental results (in terms of generalization capability and network size) of DistAl on a number of benchmark classification problems show reasonable performance compared to other learning algorithms despite its simplicity and fast learning time. Therefore, DistAl is a good candidate to various tasks that involve very large data sets (such as largescale datamining and knowledge acquisition) or that require reasonably accurate classifiers to be learned in almost real time or that use neural network learning as the inner loop of a more complex optimization process in hybrid learning systems.