X-Message-Number: 26815 From: Date: Tue, 16 Aug 2005 03:39:39 EDT Subject: Uploading technology (1.iv.0) Goldilock's choice ? Uploading technology (1.iv.0) Goldilock's choice ? Because here seems to be a low trafic these days on cryonet, I start again the brainstorming on uploading with current technology. It seems there are three possible concepts for uploading a neural system in some electronics devices: The dumb, the median and the smart. First, the dumb: In 1943, McCulloch and Pitts (1) in their threshold model demonstrated that the brain could work as a network of neurons, each with a number of positive or negative inputs on the dendrites, an algebraic linear sum at the cell body and a threshold gated firing in the axon. The thresold could be adjusted at different values for different neurons. This seems dumb, but Kleene in 1956 (2) and then Palm in 1982 (3) first hint and then fully demonstrated that such a network can solve any problem. It is a complete calculator similiar in potential power to ordinary computers. Simply, it seems more akin to the brain and best fitted to copy a neuron system. Most present day neural networks are built on that concept. The billion of neurons is not far in the future. The only problem is that biological research has advanced a lot between 1943 and now and we know that a biological neuron is far more complex that what was understood in 1943. That don't invalidate the first concept, simply up to 100,000 elementary electronics neuron would be used to simulate a single true, biological neuron. Given that conversion factor, actual neural nets are far from the biological brain power of a small animal. On the other hand, much neural nets are built from discrete components. If they was integrated at the scale of a computer microprocessor with 100 millions transistors on each chip, at least 100,000 electronics neurons could be built here. This would be equivalent to a single biological neuron. Given that the processing speed would be at least 10,000 time larger, time share would allows the simulation of 10,000 biological neuron per chip. A brain would need one million such chip, each in the $1,000 range. Not impossible, simply difficult and costly. The median solution looks at the current understanding of biological neurons and aim to build a neuromorphing electronics circuit able to copy directly all known biological properties.Neuromorphing circuits are a new concept dating back 5 years or so. This concept has been pioneered by K. Boahen as said in the 1.iii.0 message (Cryonet #26325). I'll come back on that device in another message. Suffice to say here that it must take into account information processing at presynaptict, post-synaptic, dendrite section, dendrite trees, soma and axon. This is not a simple bunch of electronics gates, it is a full microprocessor. The largest present day chips could harbor may be up to one thousandt of them. Given a time share factor near 1,000 a single chip would process one millions neurons. A brain would include 10,000 chips for a cost in the $10 millions range. The smart way was suggested by M. Soloviev : It would simulate on each processor a sub-network of biological neurons. The idea is that neurons may not be optimised and some economy can be done if we take a lot of them and work out the package as a black box. Knowing only the input and looking for the output. How could we do that ? Given that we can't test all possible input, how to guess all potential output? Worst: For cryonics, neurons are no more active, so all the potential output must be deduced from the scanned structure. The solution can come from the precomputer era of the classical physics. Assume we have a small physical system with a small number of freedom degrees or "dimensions". Everything can be computed from the newtonian dynamics resting on the formula : Force = mass x acceleration. Unfortunately, when the number of freedom degree of the system go up, even by a modest value, the computation becomes so complex that it is impossible for all practical purpose. The solution is then today to build a simulation on computer. Well, but what about the era before cheap electronics computing ? Many bright mind have taken that problem and have found new formulation of the dynamics, for example Lagrange, Hamilton, Jacobi. These analytics methods traded computing complexity against analytics one. This concept culminated with the calculus of variation, a very powerful analytics tool able to reduce, starting from lagrangian formulation of dynamics, the computing load at the price of advanced analysis. People having studied the basic logical circuits can remenber the Karnaugh diagrams a tool used to simplify a logical network. Here, the basic idea is the same, but the task is far more complex, the starting element is not the boolean algebra, it is the full real world. Neurons can be seen as dynamical systems an so can be worked out as variational systems of greath complexity. Can we take a brain column with may be 10,000 to 100,000 neurons, translate it into a variational problem and solve it, producing a small bunch of final computation? The analytic task would be daunting. On the other hand, the benefit would be enormous : A present day computer, on the order of 1,000 PC power could simulate a full human brain. The key to this concept is that we have some sotfwares such Mathematica, able to do the analytics work for us. So, what solution must we choose? The dumb one is the main track today and will continue to be so in the academic environment. It seems the hard way, but may succeed given sufficient time and money. It would be counterproductive to invest in it, this is a playground for state financed research institute or big corporation or charity grants. The median solution seems the most risk-free today with the best prospect for a continuous growth. The smart one could follow on its heel, but this is a totally new concept and we don't know what problem can be found on the road. The best strategy could be to implement the median solution step by step and at each completed step see if we can convert it to the smart one. For example, the first step could be to define a complete electronics neuron and implement it on FPGA. At the same time, a "variational neuron" would be studied. The next round would be to produce an ASIC chip able, with the time sharing technology, to simulate 1,000 neurons. The smart variational work would try to do the same job in a small software package. May be some brain area may be simpler to work out with one technology or another. For example the sound processing area seems very structured with similar circuits, it would be a prime candidate for variational compression. Olfactory domain contains random linked neurons, it seems difficult to compress such a network and the median solution may be the best here. Visual area may be intermediate, with neverthless a good hope of variational compression. (1) W. S. McCulloch and W.Pitts (1943) A logical calculusof idea immanent in neural nets. Bull.Math.Biophys. vol.5 p.115-137. (2) S.C. Kleene (1956) Representation of events in nerve nets and finite automata. In: Automata Studies, C.E. Shannon and J.McCarthy, eds. p. 3-41 Princeton Univ. Press. (3) G. Palm (1982) Neural Assemblies, Springer, Berlin. Yvan Bozzonetti. Content-Type: text/html; charset="US-ASCII" [ AUTOMATICALLY SKIPPING HTML ENCODING! ] Rate This Message: http://www.cryonet.org/cgi-bin/rate.cgi?msg=26815