I understand that neural networks model biological neurons. Each node in the network represents a neuron cell and the connections between nodes represent the connections between cells. As in nature, a neuron fires an electrical signal to connected neurons based on some kind of threshold or function that mimics such.
Recent discoveries on how the brain works reveal the importance of calcium within the cells. See http://link.springer.com/article/10.1007/BF01794675 for more information. To summarize, calcium affects the regulation, stimulation and transmission of electrical activity as well as the destruction of neurones.
From my study of neural networks, there does not seem to be a calcium equivalent. Having one would imply that the functions, connections and weights in an artificial network are configured during the training and execution process and can change over time. I understand that back-propagation is used to train the weights, but have not seen anything that trains the function nor the connections (although a zero weight could imply no connection).
Does anyone know of such a network (or training algorithm)? If so, do these networks perform better than a network that is pre-configured?