X-Message-Number: 27017 From: Date: Mon, 12 Sep 2005 15:36:57 EDT Subject: Uploading technology (1.iv.2) Strategy. Uploading technology (1.iv.2) Strategy. As Spirit at the summit of Husband hill in the martian Columbia range, we can look back at the path traveled and see the horizon on 360 . At start, there is a need for a brain scanner and a computer software to translate the picture into tomething to be implemented in an electronics system. The question is what system? Here are four basic possibilities: - An FPGA network. - An asic chip set of neuromorphic elements. - A computer or special chips running "compressed neurons" using variational lagrangian dynamics models. - A computer running a set of black boxes, one for each brain column. Here, "computer" stands for general purpose computers, it may be a single processor machine, a set of interconnected independent computers or a parallel system, for example an hypercube. It is assumed that each solution is as good as another on the theoretical ground. In practice, one may be cheaper than another or simpler to get. A strategy would mix these solutions in one way or another. For example, the FPGAs or neuromorphic chips could work on fast signals and a general computer would process the slower ones. A strategy is a path defined in taking into account a number of parameters, for example cost, security, rapidity of implementation, performances, and so on. Take for example performance, what is the best solution? It seems to be dedicated chips running variational neurons, may be at the microcolumn elementary unit. What are the drawback ? Here are some: High mathematical knowledge, complex programming, use of special chips. To be short: It is a long and costly solution. As its often the case, when these initial problems are solved, the variational neurochips cive the best result at the lowest cost because they minimise the hardware parts. So this a good choice for advanced systems in a large market. At the other side, FPGAs are the simplest solution using off the shelf parts. No special knowledge, no special chips, products readily available... On the dark side is the low capacity of current chips. A roden brain could be build this way at some cost. The human cortex would be very costly with them. A large brain seems in need of some asic (specialy built) chips mixing digital and analogic processing. The simplest option here would be a TRAC (Totally Reconfigurable Analogic Circuit)-like FPGA. Put simply, an FPGA whose elementary cells are an operational amplifier with eight selectable configurations. Given some money at start, this seems the best choice, a brain could be built ten to fifteen years from now. If the money is scarce and the time table short, a solution could be the general purpose computer running a set of black boxes. A small FPGA set would run a single neuronal column at a time and define its output. When that would be done, the column would be simulated on the computer as a black box. What must be simulated down at the synaptic level is then something as 100,000 neurons. Assume a single neuron simulated at a given instant on a chip and a time sharing factor near 10,000. Ten FPGAs would be sufficient for the project. The speed would be may be one tenth of the biological system but that is not a major problem. This simulation is not for the exterior world, it is simply for producing the output model of a black box. So here could be a strategy in three steps: A) Create a single column neurons on an FPGA set and run the results as black boxes on general purpose computers. B) Create a neuromorphic chip with TRAC technology and run directly an electronics brain. C) optimize the electronics brain with variational methods for a mass produced technology. YvanBozzonetti. Content-Type: text/html; charset="ISO-8859-1" [ AUTOMATICALLY SKIPPING HTML ENCODING! ] Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=27017