Synfire structures and cognition : a complex systems perspective

University of Technology, Sydney. Dept. of Software Engineering.NO FULL TEXT AVAILABLE. Access is restricted indefinitely. The hardcopy may be available for consultation at the UTS Library.NO FULL TEXT AVAILABLE. Access is restricted indefinitely. ----- This thesis explores the relationship between biologically realistic neural networks and cognition, from a complex systems perspective. The focus is on compositional systems of synfire chains in neocortex. A general framework is proposed for memory, learning and perception in a synfire compositional system, integrating processes across multiple time scales, wherein compositions embodying knowledge form a self-maintaining system of cooperatively co-evolving synaptic weight patterns which can be reconstructed endogenously through the coupling of wave retrieval and Hebbian weight dynamics. A network of conductance-based spiking neurons constituting a superposition of synfire chains incorporating inhibitory neurons is investigated using simulations and analysis. Through simulations, the main phases of behaviour are characterised in relation to key parameters, and cross-synchronisation of waves on cross-linked chains is studied, thereby demonstrating its strong dependence on background input: as this increases, both the speed of synchronisation and the range of temporal offsets that allow for synchronisation are reduced. Analysis of storage capacity for the superposition network is conducted using single-neuron simulations to characterise probability of firing in response to approximately synchronous excitatory inputs accompanied by background input due to crosstalk. The latter comprises random streams of excitatory and inhibitory inputs as received by cortical neurons in vivo, which for a conductance-based neuron greatly reduces the membrane time constant and raises the equilibrium potential. The resulting spurious firing rate is obtained analytically; the need to keep this under control sets the storage capacity. Optimal parameter choices within a biologically plausible range give capacities well in excess of unity, indicating that the use of conductance based neurons (compared to neurons with fixed post-synaptic potential amplitudes) gives a better trade-off between synfire propagation and control of spurious firing. A theoretical approach to the evolution of wave activity in a synfire compositional system is developed, whereby background input regulates wave activity by reducing wave survival, cross-synchronisation and wave births as net activity increases. Composite wave formation is viewed as spreading activation on a compositional landscape, with background input the control parameter for the analog of a percolation phase transition. By considering specific compositional topologies (random graph, small world and recursive compositionality) the features that emerge are critically poised effective connectivity and 'seasonal' oscillations. These constitute a cognitive cycle in which alternative composite wave 'hypotheses' proliferate during quiet conditions and then consolidate and compete during noisy conditions, with the most coherent large composite waves emerging as the winners involved in acts of cognition