Ph D abstract - Baran Curuklu
In this thesis a model of the primary visual cortex (V1) is presented.
The centerpiece of this model is an abstract hypercolumn model, derived
from the Bayesian Confidence Propagation Neural Network (BCPNN). This
model functions as a building block of the proposed laminar V1 model,
which consists of layer 4 and 2/3 components.
The V1 model is developed during exposure to visual input using
BCPNN incremental learning rule. The connectivity pattern demonstrated
by this correlation-based network model is similar to that of V1. In
both modeled cortical layers local horizontal connections are dense,
whereas long-range horizontal connections are sparse. Layer 4 local
horizontal connections are biased towards the iso-orientation domain,
whereas long-range horizontal connections are equally distributed
between all orientation domains. In contrast, both local and long-range
horizontal connections of the layer 2/3 are biased towards the
iso-orientation domains. The layer 2/3 network is axially specific as
well. Thus, this V1 model demonstrates how the recurrent connections
be self-organized and generate a cortex like connectivity pattern.
Furthermore, in both layers inhibition operates within a modeled
hypercolumn. This is in line with what is found in the V1, i.e.
inhibition is mainly local, whereas excitation extends far beyond the
inhibitory network. Observe also that neither excitation nor inhibition
dominates the network.
Based on this connectivity pattern the V1 model addresses several
response properties of the neurons, such as orientation selectivity,
contrast-invariance of orientation tuning, response saturation followed
by normalization, cross-orientation inhibition. Configuration-specific
facilitation phenomena are explained by the axially specific layer 2/3
long-range horizontal connections. It is hypothesized that spike and
burst synchronization might aid this process.
The main conclusion drawn is that it is possible to explain
connectivity as well as several response properties of the neurons by a
general V1 model, which is faithful to the known anatomy and physiology
of the neocortex. Thus, when simplicity is combined with biological
plausibility the models can give valuable insight into structure and
function of cortical circuitry.
*Keywords: *primary visual cortex, hypercolumn, cortical microcircuit,
attractor network, recurrent artificial neural network, Bayesian
confidence propagation neural network, developmental models,
intracortical connections, long-range horizontal connections,
orientation selectivity, response saturation, normalization,
contrast-invariance of orientation selectivity, configuration-specific
facilitation, summation pools