On Sun, May 07, 2006 at 09:29:51AM -0400, Ben Goertzel wrote: > However, this does not imply that in an AI, these things cannot be > done using explicit logic operations.
It's possible to build anything from NAND gates. But in practice, there are usually other constraints to implementation. Transistors and wires don't reconfigure themselves on a 3d lattice, and don't form 1:10^4 fanout on a microwatt/mm^3 energy budget. > Emulation of the human brain is only one possible route to AGI. I There are many possible routes. But in practice whoever achieves scalabe AGI first renders everything else irrelevant. There will be no close second runners, similiarly to the fact there life emerged only once on the prebiotic Earth. > agree it is a route that is ultimately sure to be feasible, but that > doesn't mean it will necessarily be the *fastest* route given our very > limited knowledge of the human brain. Is it necessary to know every single detail of mammal brain operation in order to extract constraints for a fertile seed? I don't think so. We know that we can think. Our knowledge about our CNS increases monotonically (and in fact, arguably accelerating if not outright exponentially). In contrast, we don't have any successful AI instances to study. If I had to bet money on whether biologically inspired or de novo hit region of fertility first, I would certainly bet on the former. > I agree with all this, but you have not explained by (for example) > probabilistic logic is a bad knowledge representation for > sensory-motor patterns. I think it is just as good a representation > as for instance a neural network from a pure sensorimotor point of > view' and better in terms of its interoperability with abstract > cognition. > > Do you think that, for instance, > > "high conductance along the synapses between neuron in cell assembly A > and cell assembly B" > > is somehow a superior representation to > > "P(B|A) is large" > > If so, why? In what manner? Because it's not what the wetware actually does. You picked up just one facet of the operation space. > If you explain to me what non-logic based reprsentation you prefer to > use to represent a particular piece of knowledge, I can almost surely What if you don't yet know which piece of knowledge you have to represent? And that there are no clear specs, and the requirements change constantly? > explain to you how to use probabilistic logic to express this > knowledge with equal or greater convenience, and easier > interoperability with cognition. Let's just look at a yet another isolated aspect of bio infoprocessing: using system noise to enhance computation http://diwww.epfl.ch/~gerstner/SPNM/node32.html Digital components are just analog components operated in a specific regime, and if you approach the operation limits (shrink the feature size, reduce the switching energy) you're bound to wind up with unreliable logics. Formally, it is very difficult to treat classes of circuits, classes of computations, with noise/unreliability, and shape of the mapping of input to output. Evolution has no problems treating such elements constructively. > Implicit knowledge representation such as is achieved in an attractor > neural network, can also be achieved in a network of probabilistic > logic relationships: the knowledge implicit in a set of logic > relationships consists of those relationships that can be generated > from the set via a short or otherwise simple-to-obtain inference > trajectory. > > It is quite possible to use logic expressions to interrelate raw > percepts ("pixelAt(200,150) has color blue" for instance) rather than > abstract symbolic tokens ("bird", "animal") ---- this just happens not > to be how logic has typically been used in AI systems, due to > inadequate AI designs. Do you know any AI rewriting the code base (whether at SEXPR or FPGA gate level)? How does one implement an efficient Kolmogorov predictor in FPGA gates, given just the signal sequence to predict? How would such a predictor deal with noise in the input? There must have been some recent progress in practical AI at the architecture level. Since this is not my field, I would welcome pointers. E.g. is there any trends in what DARPA Grand Challenge winners are using? -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> http://leitl.org ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
signature.asc
Description: Digital signature
