A particular group of neural network (NN) learning algorithms known as constructive algorithms (CoNN) congregates algorithms that dynamically combine two processes: (1) the definition of the NN architecture and (2) learning. Generally both processes alternate, depending on each others′ performance. During training CoNN algorithms incrementally add hidden neurons and connections to the network until some stopping criterion is satisfied. This paper describes an investigation into the semantic role played by the hidden neurons added into the NN, when learning Boolean functions. Five CoNN algorithms namely Tower, Pyramid, Tiling, Perceptron-Cascade and Shift are examined in that respect. Results show that hidden neurons represent Boolean sub-expressions whose combination represents a disjunction of prime implicants.