Principles for the Foundation of Integrated Higher Cognition

Principles for the Foundation of Integrated Higher Cognition Kai-Uwe K uhnberger (kkuehnbe@uos.de) Institute of Cognitive Science, AI Group, University of Osnabr¨uck Albrechtstr. 28, 49076 Osnabr¨uck, Germany Keywords: Integrated Cognition; Analogies; Ontologies. updates Reasoning Engine Based on Analogies Cognitive Diversity Higher cognition of humans occurs in a variety of forms, con- texts, and facets. For example, humans can perform deduc- tions, inductions, and abductions, they can solve problems in domains initially completely unknown to them, and they are able to retrieve relevant information, although there is a tremendous amount of related knowledge stored in their memories. Furthermore abilities like creativity, adaptation, learning, and reasoning with inconsistencies, by analogy, or by rough estimations are remarkable capacities of humans. In order to model such a variety of abilities compu- tationally, researchers either develop complicated architec- tures with a large number of different modules (e.g. Wang, 2006; Newell, 1990), often presupposing sophisticated con- trol mechanisms, or they try to reduce such cognitive man- ifolds to a few principles (e.g. Cassimatis, 2006; Forbus & Hinrichs, 2006). In the spirit of the latter approaches, we pro- pose to reduce the enormous variety of higher cognitive abil- ities to analogical reasoning, dynamic updates of background knowledge, and neuro-symbolic integration. Dynamically Updated Ontological Background restricts s er s sw r i e an que a qu nsw er ers ies tra i ns tra in s Neuro- Symbolic Integration Module Figure 1: A simplified architecture for integrated cognition. A rather new approach in this direction is Gust, K¨uhnberger & Geibel (2007): a semi-symbolic level is used to translate full first-order theories into a homogeneous data structure in order to train neural networks. The trained network repre- sents complex models of logical theories and was tested on benchmark problems for symbolic theorem provers. Integrated Cognition Spanning Cognitive Manifolds The integration of these modules in one architecture, although still ongoing research, could be achieved as follows (cf. Fig- ure 1): on the one hand, ontologies restrict the range of possible analogical relations computed by the analogy en- gine. On the other hand, generating new analogical relations can be used to update dynamically ontologies. The neuro- symbolic integration module learns models of input data, an- swers queries whether new analogical and ontological rela- tions should be established or not, and shows a robust behav- ior, if inconsistencies occur. Every module of the architecture learns permanently from input and computed data and inter- acts in a non-trivial way with other modules. Integrated cog- nition is the result of this interaction and could be a general basis for the variety of higher cognition, such as adaptation, non-classical forms of reasoning, and creativity. The Principles With respect to the first principle, analogical reasoning is of- ten assumed to be a candidate for creativity and many aspects of non-classical reasoning abilities. In Gust, K¨uhnberger & Schmid (2006), heuristic-driven theory projection (HDTP) is used in order to model creative analogies, to learn from a few examples, to perform non-classical inferences, or to establish non-conventional meanings of metaphors. It is therefore a framework for solving several sorts of hard problems in AI. The second principle is based on the idea that background knowledge should not be considered as a fixed conceptual- ization of the world, but as a flexible resource that is adapted, updated, and manipulated on the fly, and constantly rewrit- ten due to new input data. In Ovchinnikova & K¨uhnberger (2006), algorithms are proposed that allow an implementa- tion of dynamic changes of ontologies for text technological and semantic web applications. Although this domain is not broad enough to explain cognitive manifolds, it is a first step towards a computational model of semantic updates. The third principle is based on neuro-symbolic integra- tion techniques. By learning models of reality on the neuro- symbolic level, it is possible to avoid complicated deduction mechanisms in a cognitive architecture. Good examples for such approaches are connectionist networks trained on com- plex data structures (cf. Brown & Sun, 2000 for an overview). References Brown, A. & Sun, R. (2000). Connectionist inference models, Neural Networks Cassimatis, N. (2006). A Cognitive Substrate for Achieving Human-Level Intelligence, AI Magazine 27(2):45–56. Forbus, K. & Hinrichs, T. (2006). Companion on Cognitive Systems: A Step towards Human-Level Intelligence, AI Magazine 27(2):82–95. Gust, H., K¨uhnberger, K.-U. & Geibel, P. (2007). Learning Models of Predicate Logical Theories with Neural Networks Based on Topos Theory, to appear in P. Hitzler & B. Hammer (eds.): Perspectives of Neural-Symbolic Integration, Springer. Gust, H., K¨uhnberger, K.-U. & Schmid, U. (2006). Metaphors and Heuristic-Driven Theory Projection (HDTP). Theoretical Computer Science 354(1):98-117. Newell, A (1990). Unified Theories of Cognition, Harvard University Press, 1990. Ovchinnikova, E. & K¨uhnberger, K.-U. (2006). Aspects of Automatic Ontology Ex- tension: Adapting and Regeneralizing Dynamic Updates, in M. Orgun & T. Meyer (eds.): Advances in Ontologies 2006, Conferences in Research and Practice in In- formation Technology, vol. 72, pp 52-60. Wang, P. (2006). Rigid Flexibility: The Logic of Intelligence, Springer, 2006.