Dorian, On Wed, Sep 25, 2013 at 5:44 AM, Dorian Aur <[email protected]> wrote:
> We need to use their power of computation directly, a hybrid model > (evolving brain - digital computer) will solve many issues regarding in > vivo monitoring > I'm not sure what you are saying here. All "monitoring" already goes directly into a digital computer, if for no other reason than to clean up the signal and compute things like dV/dt and peak values that may have been obscured by limited bandwidth and sampling rate. However, there is more that remains invisible than there is that can be seen. The ONLY thing that can now be seen is the instantaneous voltage, and to my knowledge there are NO proposals to monitor anything else. This is what detoured neural networks - the belief that the ONLY thing being transmitted is a single value, the "output". Now, most AGI projects seek to perpetuate this detour, believing that everything can be contained in a "probability". Everyone seems to believe in SOME sort of back propagation, and the inability to make any sort of "simple" back propagation learn anything but orders of magnitude slower than we learn suggests that there is something more complex happening, like maybe more than one thing being propagated back. There may be other signals going forward, like some indication of plasticity to help operate the complex triage of determining WHICH neuron should change when an apparently impossible set of signals arrives somewhere. My own suspicions are that some of this may work by impedance, where some outputs are "stiffer" than others. A plastic neuron might have a higher output impedance, to be more easily affected by downstream loading. However, it still seems that whatever happens at a synapse must travel backwards to the inputs to be able to make an entire system self-organize. Until we can see and affect more, I don't see any prospect for useful hybrid computation. This present inability to interact (other than voltage) with a living neuron that is carrying multiple signals in various directions seems to me to be an even bigger technical barrier than scanning out entire connetomes, uploading and downloading, etc. The AGI folks here wave all this off, but until SOMEONE here builds SOMETHING that is capable of the same sorts of instantaneous self-adapting learning that we and every other living thing can do, AGI has no possible future. I can't see any way around this barrier, and in the years that I have been on this forum, apparently no one else has seen any way around this barrier. For those who haven't immersed themselves in this sort of code, the challenge is that learning must take place with unreliable inputs that are themselves learning and adapting. What they "learn" is often superstitious learning based on some "feature" (bug) in a signal that learning later "corrects", only to help some downstream processes while hurting others. Multiply this problem by ALL of the computations being performed doing this, and nothing works UNLESS you introduce some sort of low-pass "dampening" to keep single miss-learned computations from undoing the entire process, which destroys the instantaneous learning that is sought - or alternatively, downstream processes fall into a learning hole where they fail to learn anything that is useful. It is hard to fully appreciate this phenomenon until you have been there. At first this appears to be a simple algorithmic malfunction, until you realize that the problem presently lacks solution. >From thousands of miles away and reading postings here, people appear to be in various phases of "hacking" the above mentioned challenge. I suspect that Ben is starting to realize that there are more fundamental issues. It seems pretty obvious to me that this is from a shortfall in theory, and NOT the result of any shortfall in "hacking". I have attempted to dissect this issue and suggest possible approaches on past postings, but got no reasoned responses. It appears that to be active in AGI that it is absolutely necessary to live in a sort of La-La land of denial of this VERY basic issue. It seems to me that everyone here should *STOP* and deal with this "little detail", because AGI has nowhere to go without it. Steve ===================== > On Tue, Sep 24, 2013 at 10:29 PM, Steve Richfield < > [email protected]> wrote: > >> Back around 1970 I heard about a researcher who had a couple hundred >> pipette electrodes, all in parallel and affixed to a square plate, that was >> pushed into a brain to read a couple hundred extracellular points in the >> brain. Unfortunately, I don't remember enough about the reference to >> exhibit it here. I remember hearing about this from William Calvin. >> >> Note that the equipment didn't then exist to record and analyze this many >> parallel real-time inputs, so researchers had to switch their limited >> monitoring equipment between electrodes. >> >> This would establish the rate of progress at approximately zero, and the >> time to monitor the entire brain as infinite. >> >> Note that none of the past or present approaches to monitoring monitor >> anything but voltage. There are various ions being bidirectionally moved >> around to compute far more than can be seen by voltage alone, and these >> remain beyond our "modern" technology to observe. >> >> Further, all present approaches to monitoring KILL some percentage of the >> neurons that they seek to monitor. To scale, looking at an axon is a lot >> like monitoring YOU by stabbing you with a telephone pole sized electrode. >> Axons survive this better than people, but often all you monitor is the >> last seconds of the death of the neuron. >> >> No, for these and other reasons I reject the idea that we are making ANY >> significant progress in this area. >> >> Steve >> ===================== >> >> On Tue, Sep 24, 2013 at 11:57 AM, tintner michael < >> [email protected]> wrote: >> >>> "Out of a human brain's 100 billion neurons, researchers can presently >>> monitor only about 200 at a time. This is sort of like trying to predict a >>> presidential election by polling three people. " >>> >>> >>> http://www.psychologytoday.com/blog/world-wide-mind/201101/new-moores-law-neuroscience >>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> >> >> -- >> Full employment can be had with the stoke of a pen. Simply institute a >> six hour workday. That will easily create enough new jobs to bring back >> full employment. >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/17795807-366cfa2a> | >> Modify <https://www.listbox.com/member/?&;> Your Subscription >> <http://www.listbox.com> >> > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | > Modify<https://www.listbox.com/member/?&;>Your Subscription > <http://www.listbox.com> > -- Full employment can be had with the stoke of a pen. Simply institute a six hour workday. That will easily create enough new jobs to bring back full employment. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
