X-Message-Number: 27631 From: Date: Fri, 17 Feb 2006 15:16:39 EST Subject: Uploading (3.v.1) The gap junction neuron. Uploading (3.v.1) The gap junction neuron. Gap junction may leak potentials from one neuron body to another. This bring the second neuron up to some percents of the potential in the first. If ten percents of the somewhat 300 neurons in a microcolumn are at firing level, then they can bring nearly all other to do the same. This is why it is so difficult to separate one neuron from others in a microcolumn. For lower potentials, there may be a noticeable effect at the dendrite domain scale, such a domain is assumed to be the elementary element for the gap juncton activity. For uploading, gap junctions are of particular interest. Indeed, assume there is a "blank brain", how to load it in a person? This must be done in two steps : First, by adjusting the network geometry and second by loading the right sensibility value at each synapse. This can be done as a training using a firing threshold adjusted by the gap junction potential. So, gap junctions are the programming entries of the network. The gap junction number or kind, dont evolve rapidly, in a first system, it can even being a given property of the network. The reason to do that is that the original structure is copied from a biological configuration, an evolved one would rest on the running of the electronics system. At least for the first generations, such ystem are only approximate models. So, it may not be wise to allow too much flexibility and possibilities to evolve way out of what could be done in a biological brain. The full gap junction network must be modeled and run at the microcolumn level and from that, a "leak" potential is observed from each neuron to each other. If there are 300 neurons, this is a big 300 x 300 matrix. In fact, from it, a simplified system can be produced. Assume the maximum potential shift is 50 mV and maximum coupling is 10 percent, one neuron can't induce more than 5 mV in a nearby one. If the noise level is .1 mV, there can't be more than 50 different interaction levels, from .2 to 10 percent. This can be represented as a 6 bit number (2 at power six = 64). A local effect is then computed as sum of 300 products, each made of a neuron potential and a 6 bits percentage. This has to be done at the dendrite domain level, not the full neuron. On the other hand, nearby dendrite domain of different neurons in the same area may undergo a similar or identical gap junction effect. In fact, in a given microcolumn, the number of domains may be smaller than the number of neurons. Gap junction induced potentials have an interest beyond the neuron simulation : Because they define the firing threshold of the neuron, if this one is too high, there is few probabilities for an effective firing. So, it may a waste of time and computing ressource to simulate it precisely at this instant. A rough estimate of the activity can be deduced from nearby neurons and that may be sufficient to actualise the gap junction potential. A neuron is firing something as ten time per second, being able to predict when is a big economy. From Shannon theorem we would have to simulate it every millisecond, from this smart choice this may be reducend by a factor of 100. This could bypass the necessity to use a metaneuron level and simulate a brain directly from its elementary neurons. More precisely, the metaneuron approximation would be reserved for some particular functions not linked to conciousness, such that signal processing at low level. A single chip such the Vitex 4 SX 35 could simulate in real time more than ten millions neurons. An ASIC derivative with up to 20 millions electronics gates would reach the 300 millions neurons mark. Less than 20 would simulate a full neocortex. We have now a good idea of what could be done with current technologies. Even without ASICs, current FPGA could display a power well beyond what is found in an insect for example. A bee has 800 000 neurons, a small mammal 100 millions, the chip is not far from the second. Animals have a tremendous capacity at recognizing patterns, being visual , sound or another. Potential applications are numerous and this must be a strong incentive to implement such a neural system in FPGA. On the other hand, we may have to build a first generation brain scanner before we can wire correctly such a network. Y. Bozzonetti. Content-Type: text/html; charset="US-ASCII" [ AUTOMATICALLY SKIPPING HTML ENCODING! ] Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=27631