Spikes, synchrony, sequences and Schistocerca’s sense of smell

This thesis starts from the assumption that individual neuronal action potentials (spikes) have computational and dynamical significance. Two of the types of activity that networks of spiking neurons can engage in are sequences and synchrony. The first part of the work reviews the role spikes, sequences and synchrony play in coding, dynamics and learning in the nervous system and models of the nervous system. Models of spiking neurons, especially the spike response model (SRM), feature strongly, as do synfire chains, a form of spatiotemporal sequence. A methodology chapter deals with the problem of efficient simulation of networks of threshold-fire neurons such as integrate-and-fire (IF) neurons and SRM neurons. I show that networks of SRM neurons can be simulated with larger time steps than are required for numerical integration of equivalent networks of IF neurons. I extend an interpolation method for more accurate simulation of IF neurons to noiseless and stochastically-firing SRM neurons, and show that a network of noiseless, interpolated SRM neurons can be simulated with larger time step than the equivalent network of interpolated IF neurons. Synfire chains can be learned with a temporal learning rule and a supervised training protocol. I extend previous analyses of the speed of recall of a synfire chain by (a) explicitly including the speed at which the synfire chain was trained and (b) performing an analysis on a synfire chain comprising discrete neurons rather than starting from a continuum approximation. I conclude that synfire chains can be recalled much faster than the speed at which they were trained. The final part of the thesis is a modelling study of a system that incorporates spikes, sequences and synchrony, the olfactory system of the locust Schistocerca gregaria. Building the model leads to the problem of setting the mean weights in a network of reciprocally-connected excitatory and inhibitory spiking neurons on the basis of the desired oscillation time period and phase lead of the excitatory neurons over the inhibitory neurons. The solution to this problem includes a generalisation of a locking theorem from the literature to certain types of networks with nonuniform connections. Finally, I integrate the theoretically-determined mean weights into the locust model and present simulation results from it.