I'm using the terms learning and development in strictly-defined ways, since the concepts are often fuzzy, and the distinctions are often blurred.

A synapse is the place where two neurons interact (or a single neuron interacts with itself). It's actually a narrow gap, called a synaptic cleft. I won't go into all the details of neurotransmitters and how they work, but basically the synapse is the place where information is stored. A change in synaptic strength is a result of learning. In a neural network model, synaptic strength is represented by a connection weight. So learning is the change in connection weights.

New neurons are generated, and new connections are formed as the neuron's axons (outgoing cables) and dendrites (incoming cables) branch out. Typically there's an overproduction of cells and connections early in human development, and a great deal of pruning (selective dying off of cells and connections) as time goes on. The formation of the brain is analogous to sculpting, and it's almost as if nature produces an overabundance of raw material, like a huge block of marble, and then carves away the unnecessary bits over time to stabilize on a mature brain. When I talk about development I'm referring specifically to cell growth and death, and connection formation and pruning.

In my model, learning and development are both a result of similar mechanisms, primarily something called Hebbian dynamics, named after Donald Hebb.

It's a pretty simple idea, really. Basically, if two neurons are firing strongly at the same time, the synaptic weight between them is strengthened. This idea is often summarized as: Neurons that fire together wire together. The assumption is that if two neurons are firing around the same time, whatever stimulated them must be correlated. Unlike some other artificial neural network training methods, simulated Hebbian learning is biologically plausible, and there's strong evidence from neuroscience that it is a fundamental mechanism of learning. The phenomenon has been repeatedly observed in a number of species, and the process by which synapses are strengthened due to pre- and post-synaptic activation is called long-term potentiation.

Conversely, there is something called long-term depression, which is a decrease in the sensitivity of a synapse, basically weakening its effect. There is some evidence that this is the result of decorrelated activity. Basically, if one neuron is firing strongly, and the other one isn't, the connection between them is weakened.

This is basically how connection weights are strengthened and weakened in my model, in other words, how learning occurs.

As for development, new connections are formed as a result of correlated activity. If two unconnected neurons are both firing strongly at the same time, there is an increased probability that they will become connected. And conversely, if two connected neurons are firing at different times, their connection weight will tend to be weakened. The probability that an existing connection will be pruned is a function of how strong the connection is. So every once in a while, connections will die off, and those that are very weak will have a higher chance of being pruned.

These choices are all based on evidence from neuroscience, that large-scale wiring patterns are determined genetically, while the fine-scaled wiring in the neocortex is molded by experience. We know that rats raised in impoverished surroundings have much sparser neural connectivity than rats raised in stimulating environments (with lots of toys and brightly colored cages). My model will reflect these features.

I think maybe that'll wrap up my discussion of the neocortex and how I plan to model it. If anybody has any questions, let me know. I think I'll go ahead and compile links to all these entries in a single post, for ease of access, and I'll add to it in the future if anything changes or if I have more details to add.