Spiking models of local neocortical circuits.
The 'local circuits' of the mammalian neocortex are defined within cortical columns less than
1mm across. There is mounting evidence that these fine-scale neural networks are an important
organisational and functional unit. However, detailed study of the local circuitry is hampered
by considerable technical difficulties; computer-based modelling therefore offers an important
approach to understanding their basic properties.
In this thesis, computer simulations are used to examine some of the fundamental questions
associated with this class of neural network. We describe a spiking-network model which is inspired
by anatomical and physiological study of local neocortical circuitry. Small, heterogeneous
circuits are constructed from regular firing, rhythmic bursting and fast spiking neurons. These
cells interact strongly through dynamic connections; synapses exhibit facilitation, depression, or a
hybrid form of non-associative plasticity. A fast, asymmetric Hebbian process is also examined as
a model of the 'Malsburg synapse', and is implemented in parallel with the non-associative fonns
Exploration proceeds in three stages using a bottom-up methodology. First, we investigate the
dynamical repertoire of the individual classes of circuit; the significance of architectural variation
between circuits is addressed, and we examine the influence of fast adaptive processes in shaping
network dynamics. Guided by available experimental data, circuits are connected in the second
stage to create larger architectures; these are used to study interactions between the circuits. In the
third stage, the inhibitory circuits implement surround and feedback inhibition, and local circuit
effects of these two contrasting models are explored.
Simulation results offer novel links between disparate experimental data. They also indicate
how variation in the architecture of particular local circuits, and the different classes of connection
between these circuits, might have functional pertinence. More generally, our findings suggest
how this style of network may support a highly flexible, dynamically configurable computational