Page - 100 - in Limina - Grazer theologische Perspektiven, Volume 3:2

100 | www.limina-graz.eu Sara Lumbreras and Lluis Oviedo | Belief networks as complex systems In belief systems, concepts are at least partly comprehended in their re- lationships to others, so they excel at generalization. This is remarkably important in an evolutionary sense, given that humans live in a changing environment where no two situations are precisely alike. Another interesting property of belief systems that distinguishes them from the architectures of machine learning is that they provide a filter for experience. This filter would have a stronger effect the more ambiguous the situation. Many authors have explored the bidirectional relationship be- tween belief and evidence (Fryer Jr/Harms/Jackson 2019). The programmer defines the examples that will be received by an ANN, for instance as a grid of shades that encodes an image digitally. However, the categories in our belief systems define what features of reality are relevant. All experience gets filtered through the belief system, and its information is used to dynamically update the belief system. The change can happen in two ways. The first one corresponds to be- lief emergence, when a pattern of relationships between other concepts is found to appear repeatedly or in a significant manner. Belief emergence is one of the key phenomena in this context and has been subject to extensive study (Keil 1991). Belief emergence can only be linked to machine pattern recognition in a superficial way because, as explained above, animal pattern recognition is based on concepts and is more robust to generalization. When a pattern is sufficiently important, we create a category and usually give it a name. Then, the relationships between the previously held and the new concept are described (similarity, difference, proximity, etc.). Probably, only the concepts that seem to be close in terms of similarity or proximity are au- tomatically (unconsciously) scanned for their relationship to the new con- cept. It is plausible that, in a first phase, relationships are proposed based on imperfect memory and are subsequently tested for validation in practical experiments. The human brain seems to be especially adapted to detecting subpatterns that help it to generalize. A particularly enlightening example is the Doman method for reading (Doman/Doman 1994). In this method, children are presented with cards that spell out certain words. Their care- giver is supposed to read the whole words to the toddler repeatedly. After enough practice, children not only recognize the words on the cards, but Belief systems provide a filter for experience. Its categories define what features of reality are relevant.