Artificial Consciousness/Neural Correlates/Synaptic Models/Pre-Synaptic rule

Pre-Synaptic Rule for Weight Adjustment

Even though neural network implementations have developed complex feedforward and feedback networks, based on Universal Learning rules such as the Delta rule, which in turn is based on an error measurement that only makes sense, if you are training to a particular goal, there has always been the criticism that Neurons do not actually work that way, there is little evidence of a global learning rule, instead each cell in the network applies it's own set of rules. More modern Neural Networks tend to have at least two learning rules, a Pre-synaptic Learning Rule, and a Post-Synaptic learning rule.

What this means is that the actual rules by which weights are adjusted, depend both on the state of the pre-synaptic buds that the cell uses to output, and the state of the signals coming in at the synapses. Somehow these states which are situated often at opposite ends of the neuron, both impact on how often weights are increased. Ignoring the Post-Synaptic Learning rule for now, Let us look at the Pre-Synaptic Learning rule, to figure out what is happening that allows the state of the outputs of the neuron, at usually the Axon tips, to influence the inputs of the Neurons at the dendritic tips on the other end of the neuron.

My first clue as to what is going in, was found in Memories Voice by Daniel L. Alkon. In this book, he documents experiments done on snails, in an attempt to decipher the neurochemical basis of memory. One aspect he documents, is the peculiar actions of a specific chemical that seems to be generated as a result of firing of the neuron at the pre-synaptic bud, and travels retrograde up the axon, to the soma of the cell where it gets involved in protein regulation. The proteins differentially formed as a result, seem to diffuse through the cell but concentrate over time, in the dendritic processes where the post-synaptic sensitive patches can be found.

As an interesting tidbit, Dr. Alkon notes that the proteins involved, once sequestered at the synapses, seem to outlast the free-protein in the rest of the cell. This suggests either replacement, over time, or some chemical change that makes them less easily digested.

From a functional point of view, what we are looking at is feedback from the firing of the cell, somehow getting involved in the presence of protein tags at the synapse level. Another way of looking at this is a positive feedback loop that reinforces signals that act to reinforce the output of the Neuron. If only active neurons sequester the protein tags, then the availability of the tags from a previous firing of the cell, will act to reinforce the tagging of the synapses that are the most likely to have contributed to the firing in the first place. If this increases the number of ion channels in the synapse somehow, then the synapse gets stronger.

However this points out the idea that not every tag, is ephemeral, these tags can be shown to survive for weeks after the neuron has fired. Long after they would normally have been digested by the relatively hostile nature of cellular plasma. This suggests the concept of a Survivable Tag, a Tag that is either sequestered in a way that protects it from digestion, or that is replaced before it can be digested away, possible in the membrane replacement mechanism. Some scientists have suggested that in fact survivable tags, contain docking stations for ephemeral tags, so that they increase the chance that an ephemeral tag will be sequestered in the synapse. The ephemeral tags thus sequestered still get consumed at the time of synapse regeneration, but the survivable tags might get rebuilt with the rest of the synapse. The net effect of the docking station is to increase the leverage by which the ion channels can be increased.

Although Dr. Alkon was not sure of the mechanism, he noted that neurons that nearly fired but didn't, still produced his chemical, but somehow the result was not an increase in the protein output, but a decrease instead, this suggests that there might be an Unlearning component to Neural Function that reduces the number of sequestered protein tags, when the cell is inadequately stimulated and doesn't quite fire.

← Synaptic Gap Model

Artificial Consciousness/Neural Correlates/Synaptic Models

Post-Synaptic rule →