Fig. 5

Architecture of a cascade correlation network. As shown in this Figure, a Cascade Correlation network comprises an input layer, one or more hidden layers, and an output layer. The cascade architecture is characterized by its adaptive growth, where hidden neurons are introduced sequentially as the training progresses. Unlike traditional neural networks, where all hidden neurons are preset at the onset of training, in Cascade Correlation networks, once a neuron is added to the hidden layer, its incoming weights become frozen, ensuring that its pattern of activity is not modified in subsequent iterations. This dynamic architecture aids in the enhancement of the network's capability and adaptability. The second foundational principle revolves around the learning process. When introducing a new neuron to the hidden layer, the training algorithm specifically aims to maximize the correlation between the output of this newly added component and the residual error of the network. This approach ensures that each added neuron makes a significant contribution towards correcting the overall network error, leading to efficient and effective training. This combination of dynamic architecture and specialized learning makes the Cascade Correlation network a powerful tool for various tasks. Source: Abdou, et al. (2016), pp. 91–92, Modified