Page - 99 - in Limina - Grazer theologische Perspektiven, Volume 3:2

99 | www.limina-graz.eu Sara Lumbreras and Lluis Oviedo | Belief networks as complex systems More importantly, belief systems have an internal structure that has not yet been reflected in the workings of ANNs, or machine learning in general. The most important difference we would like to highlight is that belief sys- tems are built as a hierarchy of concepts or categories. A concept is a pat- tern that is comprehended, that has meaning in the sense of understanding the concepts that relate to it in the hierarchy and that are either at its same level, correspond to a partial aspect of it or correspond to a generalization of itself. The relationships between the concept and other concepts in the network can be known partially or in full and can be subject to ambiguity. Powerful concepts have a simple description and can be applied to a large number of instances, while weaker concepts are more convoluted or am- biguous. Some authors have postulated how belief emergence and change maximizes explanatory power or minimizes cognitive dissonance. This parsimonious quality of belief systems has a key property that has not been noted previously, at least to the best of our knowledge: they are robust with respect to generalization. Any machine-learning engineer working with ANNs to solve classification or forecasting problems knows that their main danger is overfitting (Hawk- ins 2004). Overfitting happens when the ANN learns the examples too well and is not able to generalize. For instance, if most of the pictures of cars we show our network correspond to vehicles photographed on a sunny day, the algorithm might have trouble recognizing a car when the surroundings are dark, as it does not recognize that lighting is not a relevant feature when identifying a car. ANNs (and any other machine-learning technique) do not understand what a car is; they are not able to discern when a characteristic is important and when it is not. They only infer this from the examples they are given. Overfitting is an issue that has to do both with the examples we show them and with the training methodology, for instance taking care to select a representative enough set of instances. The fundamental trouble with machine pattern recognition in general (not only ANNs) is that al- though the pattern might be recognized correctly, it is not comprehended. The pattern of a car, for a computer-recognized pattern, might be a com- plex combination of light and dark spots on an image. The network does not need to recognize the wheels or the engine to identify a car, so when the situations vary slightly (such as in the lighting example) they might lead to errors. Belief systems are robust with respect to generalization.