Abstract
The connections between cortical neurons change throughout an animal’s life, so that for humans and other vertebrates, the pattern of their neural connections is not defined at birth. In this chapter we simulate the impact of such reconfiguring of neural connections on the circuit’s topology and on the distribution of connection strengths. Because one neuron cannot respond to the activity of another neuron before a connection is made, we assume that neurons randomly sample their local neighborhood when initially making contacts with new cells. Such random formation of connections can lead to the observed nonrandom structure of cortical circuits if the loss of connections is nonrandom. In particular, we demonstrate that if a Hebbian-like plasticity mechanism acts to maintain connections between coactive neurons—while connections without such maintenance are gradually lost—then many of the observed nonrandom features of cortical circuits arise. We find that a randomly connected recurrent circuit can be trained to produce strong selectivity to specific pairs of inputs that it receives so long as the appropriate combination of rules for changes in synaptic weights is used and that such a circuit can produce the responses needed for good performance when trained in a linearly nonseparable task. Inclusion of structural plasticity has little impact on performance but does account for the formation of highly interconnected clusters of neurons responsive to the same stimulus. Moreover, perhaps surprisingly, we find that within such a recurrent circuit, that structural plasticity is necessary to produce the observed unimodal distribution of synaptic strengths.