A deterministic dynamic associative memory (DDAM) model for concept space representation

This dissertation aims at the general goal of solving the problem of representing and processing information on conceptual principles, in an unsupervised, human-like manner, and using existing computational methods. Given this very general context, the need for intelligent applications that meet the complexity and sensitivity requirements of Medical Informatics is postulated in what is referred to as "the axiom of medical information systems." The reformulation of the axiom that "medical information systems must be, at the same time, usable and useful" leads naturally to the identification of more immediate, achievable objectives in the form of context dependent information processing and case-based reasoning research on memory models capable of unsupervised representation and processing of information, in a similarity-based manner. Further, the unification of these objectives is proposed in the form of the general problem of managing associative concept representation spaces characterized by four fundamental properties: high dimensionality, sparseness, dynamicity and similarity based organization. The thesis of this dissertation is that the solution to this problem can be approached in the most appropriate way by memory models that specifically address each and every one of the four fundamental properties. The support for the thesis is twofold and comprises theoretical accounts which lead naturally to the definition of a memory model, the deterministic dynamic associative memory model (DDAM) which is based on the existing mathematical structure of partial order set. The model is first introduced informally by means of examples and depictions that speak for its usability. Further the formal description of the DDAM model and learning algorithms is achieved using existing fundamental concepts of set theory and combinatorics. Finally, the DDAM model is evaluated and compared with existing approaches in a series of experiments and simulations that demonstrate usefulness comparable or superior to existing approaches.