The Principles of Self-Organization of Memories in Neural Networks for Generating and Performing Cognitive Strategies

Higher-order animals exhibit the remarkable ability to dynamically adapt to a changing environment. On the neuronal level, they have to form mental representations of specific stimuli, so-called memories. Furthermore, they abstract and arrange multiple contextrelated memories into a corresponding network that can also be adapted by changes in the environment. Such adaptive networks of interconnected memories are termed schemata and construct the mental representation guiding behavior. Considering two interconnected memories within a schema, we can define three different forms of functional organizations of memories dependent on the ability of the memories to either excite or inhibit each other: Two memories can either mutually excite each other, i.e. form an association, mutually inhibit each other i.e. form a discrimination or build up an asymmetric organization, where one memory excites and the other inhibits its interconnected memory, i.e. form a sequence. In order to adapt schema to external stimuli, all of these functional organizations must emerge from the same underlying neuronal mechanism. Experimental, computational and theoretical studies have shown that the underlying neuronal mechanism forming memory representations is activity-dependent synaptic plasticity. This mechanism leads to the formation of strongly interconnected groups of neurons, so-called cell assemblies, which decode memories. However, whether the same synaptic plasticity mechanism can account for the formation of large networks of memories is still unknown. In this thesis, we derive a theoretical model of interacting neuronal populations that enables to analytically study different synaptic plasticity mechanisms with respect to their ability to form all three functional organizations of memories. Two specific excitatory synaptic plasticity mechanisms, correlation-based and homeostatic plasticity, have already successfully been used to form individual cell assemblies in neuronal networks. Nevertheless, our analysis reveals that these two plasticity mechanisms are not sufficient to implement all different forms of functional memory organizations such that further mechanisms are necessary. In this thesis, three distinct strategies are proposed that enable the formation of diverse networks of memories. The first approach is to add a further excitatory synaptic plasticity mechanism based on the causality of neuronal firing, in particular, calculating the difference of pre- and postsynaptic neuronal activities. The second strategy is to allow for inhomogeneities in the time scale of the homeostatic synaptic plasticity mechanism, serving the consolidation of individual memories. The third solution is accomplished by inhibitory synaptic plasticity in addition to correlationbased and homeostatic excitatory synaptic plasticity. However, these three distinct implementations of synaptic plasticity mechanisms are capable to enable the input-dependent formation of all three functional organizations of memories. Therefore, implementing these strategies yield complex adaptive networks of memories, hence, enabling behavior. Finally, we strongly advocate that these synaptic plasticity mechanism can be used in an dynamically input-dependent manner to compute any algorithm that is complete with respect to structured program theory. Thus, the synaptic plasticity mechanisms proposed in this thesis could be extremely useful for technical and computational applications