Feed-Forward Neural Networks

© Fraunhofer CSP
Schematic representation of a feed-forward network consisting of an input layer, two hidden layers and an output neuron. (Source: http://cs231n.github.io/neural-networks-1/)

We develop so-called feed-forward networks for solving static problems. Such static problems can be for example an image recognition problem or the interaction of process parameters in a production. Thus, this kind of neural network is basically suitable for a process optimization and simulation in so-called  digital "ab inito" experiments. Such feed-forward networks are typically based on fully interconnected neurons consisting of several layers, so-called hidden layers. Image recognition assigns an input neuron to each pixel, whereas process optimization assigns an input neuron to each process parameter. After the training of such a network, it can be used to simulate any variation of input factors and their influence to the output. This enables a digitalized product and process development in combination with low costs.