Seite - 96 - in Limina - Grazer theologische Perspektiven, Band 3:2

96 | www.limina-graz.eu Sara Lumbreras and Lluis Oviedo | Belief networks as complex systems or prediction. The programmer does not input any specific rules into the network. On the contrary, the system “learns” by means of example. For instance, an ANN can be trained to identify pictures of cars, without know- ing that cars have four wheels, seats, or a trunk. Instead, the network re- ceives pictures that have been labelled as “car” or “not a car”, and infers what the underlying characteristics of automobiles are. This is what we call supervised learning, as the system receives items that have been correctly classified. In other words: the network recognizes the pattern that appears in the examples it receives. ANNs have been remark- ably successful in performing difficult tasks such as computer vision, speed recognition, machine translation and medical diagnosis. All these appli- cations correspond to classification problems, where we need to identify which set of categories a new observation belongs to. There is a second, no less important type of problem known as forecasting. In forecasting, the network detects the patterns that underlie the time-dependent evolution of a variable and predict their unfolding. ANNs have also excelled at this task. There are many different flavours of networks, which have been proven to have varying strengths. Normally, these networks are structured in several layers, where some of the units are in direct contact with the input they receive, others constitute the output and the remaining ones stay hidden. Each of the neurons receives an input from the set of neurones that are connected to it and it uses this input to generate a single output by means of a relatively simple function, usually just a linear combination of the inputs and some weights. This linear combination is passed through an activation function that maps it into the interval [0,1]. The hyperbolic tangent or the sigmoidal function are some of the main activation functions used in ANNs. These simple calculations provide the framework for the ANN. The weights that will define each neuron are calculated by the application of what is known as a training algorithm. All training algorithms start by allocating random starting weights to the network that will be progressively adjusted taking into account the errors that they create in the outcome. Backpropa- gation is the method that calculates how a current error is related to each of the weights and how they should be adjusted. It is efficient and can be ex- ecuted with short computation times, so that it is possible to quickly obtain The programmer does not input specific rules into the network. The system “learns” by means of example.