Mike, On Tue, Oct 2, 2012 at 8:35 AM, Mike Tintner <[email protected]>wrote:
> Steve, > > Just to throw something in – not v. coherently, but the general principle, > I believe, is important. > I'll answer just as incoherently... > > S: I suspect that once an understanding emerges, that there will be > several parameters that come out of this process, e.g. one of which > determines the synaptic efficacy, another determines differentiation, etc. > Perhaps the reason that it isn't obvious is because it is a group of > methods rather than a single method? > > What seems to me to characterise both your and others’ assumptions here re > brain processing is *rigidity* – you’re assuming the brain (and body) are > rigid, algorithmic processors. Not so. > Agreed - Not so. I also believe that connections change MUCH more than most people now think they do. I suspect that much of what is now believed to be "memory" is really reconfiguration. > > How do you recognize a hand, handling something? It could be many > different shapes. > This appears to be an emergent property of ???. > > You don’t need to process it rigidly. If the visual shape your brain > registers is possibly handlike, the brain will link it up to your actual > hand – and if your hand can assume that shape then the brain can accept it > as a hand. > The BIG question is just how such a "link" might work. > > Ditto with any physical body you see in some strange form/position. If > your body can assume/ contort into that form, then it could be the > dog/cat/whatever it first seems to be. > > By processing via the body as an interpreter, you gain immense fludiity of > processing. > You are preaching to the choir here. > > And perhaps in some strange way, wh. I am far too ignorant to articulate, > the very networks of the brain are themselves fluid and highly malleable > like the external body and limbs they direct – and this is central to > processing. We know after all, – don’t we? – that brain areas can radically > reconfigure themselves for radically new functions. > Yes. > > Call this “corporate analogical processing”. > The really interesting thing I found is that by merely representing things as dP/dt (rate of change), the same old learning methods now works with temporal learning, without need for ANY other changes. Now, does real-world "learning" work like machine learning (ML), or does it work quite differently, e.g. abandoning connections that don't work well in favor of new connections that seem to work better, and tailoring the connections for the data? I wonder what the OTHER possibilities might be. You seem to be the leading "fluid thinker" pundit here. It would be interesting to try to list ALL of the possibilities we can think of as to how such a thing might conceivably work, so we can start picking through them via the Scientific Method. The main problem with fluid reconfiguration is that everyone has presumed it to be too slow to work in the real world. I doubt that this is really true, because connections that have already been made could be changed VERY quickly (a fraction of a second). Growing new connections would take longer (hours) - but there would be warning that they might be needed, e.g. because a neuron can't find anything "useful" to do. Hence, new connections could be ready and waiting when needed. It would seem important to quickly abandon useless connections, to make room for other new connections. Of course, fluidity is VERY simple in a computer, as a connection is simply a subscript in an array that is easily changed. Handling the real-time wave-shape issues is MUCH more difficult, probably requiring a real world in which to learn. There is plenty of evidence (but NO proof) that there is SOME sort of biochemical coding that controls which neurons are prospective targets of connection. I have suspected that this represents what we now call "dimensional analysis" to avoid wasting space trying to relate unrelatable things. How am I doing here? Are we singing the same tune yet? Steve ================== > *From:* Steve Richfield <[email protected]> > *Sent:* Tuesday, October 02, 2012 4:19 PM > *To:* AGI <[email protected]> > *Subject:* Re: [agi] I just bought a GP-6 analog computer... > > Mike, > > I have programmed a variety of patch-wire computers, including: > > 1. Analog computers like the TR-20. > 2. Electromechanical accounting machines, like the IBM-407. With a ~40Hz > clock, these were fast enough to do the billings for large utility > companies. > 3. Hybrid digital computers like the IBM-650, where the I/O formatting > was patch wired, while the computing was done with conventional sequential > programming (if you consider bi-quinary arithmetic to be "conventional"). > > Once you get your brain wrapped around doing things in parallel, patch > wired systems become MORE intuitive than sequential programming. There are > some subtle new problems, e.g. you can't relate two things together unless > they arrive to where they are to be related at exactly the SAME time. > > Continuing... > > On Tue, Oct 2, 2012 at 2:10 AM, Mike Tintner <[email protected]>wrote: > >> Well – as a non-techie – I like the idea of just “changing a few >> wires” – could that be what the brain is doing in its processing? >> > > That is my own best guess. > > >> And can you adapt this approach to visual processing – >> filtering/reducing the visual signal in successive stages, so it becomes an >> outline, and finally abstract? >> > > I don't see why not. The ***BIG*** challenge to making this (and all other > things AGI) work is in figuring out how connections are "figured out" and > established by individual neurons that can NOT see the whole picture, have > no "concept" of "engineering", etc. > > I suspect that there is some relatively (to what? I don't know) simple > algorithm for establishing and adjusting connections. Until this is figured > out, all efforts to build an AGI are a complete waste of time, because > everything will be discarded once the system can wire itself as needed. > > Note that "connections" may be rather complex, e.g. involve integration or > differentiation and who knows what else. I suspect that once an > understanding emerges, that there will be several parameters that come out > of this process, e.g. one of which determines the synaptic efficacy, > another determines differentiation, etc. Perhaps the reason that it isn't > obvious is because it is a group of methods rather than a single method? > > *Story:* In ~1966 while working on maintaining the Direct Coupled IBM > 7094/7040 mainframe system at the University of Washington, we were > apparently down to the very LAST bug in the operating system, which was > what we now call a "heizenbug" (randomly occurring). I called an extended > lunch meeting NOT to discuss the bug, which had already been "discussed to > death", but rather to discuss how ANY bug could possibly elude our very > expert staff. Over the course of ~2 hours various people advanced > possibilities, most of which were quickly shot down by someone else. > Eventually a viable "proposal" emerged, where something random like an I/O > error could occur in flawed non-resident code, only to have the code > discarded and overlaid by the time the entire system crashed, leaving no > trace in the memory dump of what went wrong. The bug was then found in ~2 > hours, before the end of that day, and was exactly where it was > hypothesized to be. > > *I think it possible that a similar discussion might be fruitful here on > the AGI forum. How is it possible that an apparently simple wiring > methodology has eluded generations of really smart people? > * > I'll start the discussion out with my own pet theory: I think that neurons > work with the real-time shapes of waveforms, precise temporal > relationships, etc., and NOT (much) with the "logic" of things as everyone > has previously presumed. This is implicit in my noting that real-world > neurons appear to be working with dP/dt signals rather than simple > probabilities, and switching to dP/dt representation implements temporal > learning with NO additional functionality needed. > > One thing my glaucoma cure has shown is that everything is driven by > utilization. By putting a little masking tape on one lens of my glasses > (now replaced with cute little a simulated diamond), I have not only > brought neurons in the OTHER eye back to life, but increased the SIZES of > the arteries and veins feeding them. > > Another thing: Vision is not one big homogeneous array. Instead, there is > a tiny high-resolution center (the fovea), another area extending ~20 > degrees from center that sees larger details, and the rest of the visual > field is motion-sensitive peripheral vision. I suspect that there is good > reason for chopping things up this way, e.g. really complex vision probably > doesn't work near boundaries. > > Steve > ================ > >> *From:* Steve Richfield <[email protected]> >> *Sent:* Monday, October 01, 2012 10:22 PM >> *To:* AGI <[email protected]> >> *Subject:* [agi] I just bought a GP-6 analog computer... >> >> Hi all, >> >> I was about to start work on a signal processing project to transform >> sounds that are above the range of hearing into like sounds that are within >> the range of hearing. This would serve as a hearing aid, but MUCH better >> because ordinary hearing aids CAUSE hearing damage by pelting you with >> amplified sound. The final product would be something you wear or keep in >> your shirt pocket, that adds to sounds that you hear directly with your >> ears. >> >> So, where could I find a computer that is perfectly adapted to audio >> frequency signal processing. OF COURSE, an antique analog computer, where >> with a few wires I can change things MUCH faster than you could ever >> re-engineer signal processing code. Further, there is no sampling or >> quantization noise, or other artifacts of digital implementations. >> >> This computer is a gorgeous piece of antiquity, highlighted by its very >> retro Nixie tube voltmeter display. >> >> Technology marches on - but backwards. >> >> Any thoughts? >> >> Steve >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/6952829-59a2eca5> | >> 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> > <https://www.listbox.com/member/archive/rss/303/6952829-59a2eca5> | > 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-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
