Does your adaptive control concept have anything to do with ACT-R (Adaptive Control of Thought - Rational) ? Kindly advise. ~PM
Date: Mon, 4 May 2015 16:34:38 +1000 Subject: Re: [agi] AGI as adaptive control From: [email protected] To: [email protected] On Mon, May 4, 2015 at 1:41 AM, Steve Richfield <[email protected]> wrote: Colin, WOW, there is more than one of us!!! This means we can actually have a conversation about this stuff. I will now sprinkle some comments in with your reply to get this conversation going. I suggest addressing the several issues one-at-a-time in separate threads. 3 deep is enough!This is a massive missive posting. On Sun, May 3, 2015 at 12:25 AM, Colin Hales <[email protected]> wrote: On Sat, May 2, 2015 at 2:50 AM, Steve Richfield <[email protected]> wrote: Jim, Again, I think I see the POV to solve this. All animals, from single cells to us, are fundamentally adaptive process control systems. We use our intelligence to live better and more reliably, procreate, etc., much as single-celled animals, only with MUCH richer functionality. Everything fits this hierarchy of function leading to intelligence. Then, people like those on this forum start by ignoring this and trying to create intelligence from whole cloth. This may be possible, but there is NO existence proof for this, no data to guide the effort, etc. In short, there is NO reason to expect a whole-cloth approach to work anytime during the next century (or two). However, some of the mathematics of adaptive process control is known, and I suspect the rest wouldn't be all that tough - if only SOMEONE were working on it. Erm.... guys. This would be me. I am working on it. For well over a decade now. Cognition and intelligence is implemented as an adaptive control system replicating, inorganically, the natural original called the human (mammal) nervous system. I simply replicate it inorganically. Tough job but I am getting there. There appears to be some confusion between form and function. A computation is a computation regardless of how it is performed, e.g. EM fields, electrolytic tanks, analog computation, electromechanically, etc. "The differences between electrical, chemical, and mechanical processes disappear when the scale becomes small enough" approximate quote by John von Neumann at a neuroscience conference. There's no programming. I presume you mean high-level programming. Of course there would be SOMETHING (firmware, wetware) at a low level to make things work. No software. No software. I mean it. Like there's no software or computing in a brain. I get to configure adaptation parameters. That's it. Everything else is emergent. Like the brain. And when we observe it .... it will appear to be doing computations just like the brain. The apparent computations are merely emergent regularities in its operation that can be described as computation. Not to be confused with a computer computing an abstraction of the observe/documented regularities. Of course the sticking point here will be comprehending how these two things differ and is central to my approach. Von Neumann was 100% right but, with respect, 100% irrelevant. If you want to study X you can compute abstractions of it to death. However, if you want to actually perform the role X performs in the natural context (build an X), and there is essential physics involved, then no essential physics = no X. Take flight. No amount of simulating flight physics actually flies. In the case of a robot then no amount of simulation of an arm can be an arm (that actually picks up the coffee cup for real). Normally we consider the essential physics takes over in the embodiment (input/output). The abstract model pulls the strings of the real world embodiment through the essential physics of arms and legs, motors and transducers. E.g. Like wings on an aircraft are not optional. E.g. 'software radio' measures an antenna - real essential physics. No EM field, no radio. No amount of software changes this. So what happened with me? I realised that the essential physics of the brain may extend deeper into the tissue than mere peripheral (affect/effect, motor/sensory) hardware. Or, alternatively, you can imagine that there is deeper physics that has the same input/output role that can't be simulated in practice. This 'deeper I/O'? For me: the EM field system produced by the brain. The biggest single entity in the organ. Indeed in any organ. It is a single entity (by vector field superposition) being blasted out the scalp (magnetic and electric). I say you can't simulate the fields because they have content that you cant simulate: Information from the outside world that is physical/causally only accessible via the field system. And that it has a real physical influence in the form of a 'virtual' feedback loop (physically via the Lorentz force) that is only accessible by BEING the fields. Which is what we do as humans. It is what my robots will do. BE the fields like we do. This is where the conversation usually stops. Eyes glaze over. I get labelled a space cadet and get sidelined. The thing is that to prove my position you actually build it and do it. The flight analogy: Put the computed flight physics and the replicated bird physics (a plane) on the tarmac and see which one flies. I want to actually do this for the brain. The alternative is to assume there is no essential physics beyond peripherals and then never do the science to prove it. So how I see the AGI milieu is as a massive >50 year experiment that this assumption (the lack of any essential physics deep in the brain) applies. Not by actually physically implementing and testing, but by intuition .... in assuming that there's no essential physics deep in the physical makeup of the brain, and computing and computing and computing and computing and never realising that the wrong experiment is being carried out. This is more than a technical issue. This is science-cultural. I wrote a book on this. It's the real problem of AGI. The technical solution is actually straightforward. All we have to do is resume doing actual science of the pre-computer-age kind. See above. Just radically adaptively nested looping processes. In control strategy terms it is a non-stationary system (architecture itself is adaptive). Control loops come into existence and bifurcate and vanish adaptively. The architecture commences at the level of single ion channels I hadn't really thought about individual channels, but of course you are right. and nest at multiple levels that then appear in tissue as neurons doing what they do, For the rest of our viewing audience, there is a VAST chasm between what is commonly taught about neurons, and what has actually been observed in the laboratory, but never captured in a reproducible form. The belief that if it cannot be reproduced than it is not "science" has effectively destroyed the communications channel between neuroscience labs and AGI people. For example, only a tiny fraction of neurons (apparently those with long axons) actually produce spikes - the rest appear to compute and communicate in a continuously analog form. A big yes to this. This is the low frequency 'electrotonic' control system such as that in invertebrates. But more important for me are sub-threshold oscillations. In my approach these also create a structured EM field interference pattern without requiring any action potentials and that have a potential perceptual/regulation/adaptation role to be discovered. There is a vast undiscovered country to be explored here. Note also that as time goes on all empirical work zeroes ever more closely in on the EM field (usually called the 'local field potential') as the most direct 'correlate' of consciousness. but need not appear like this in the inorganic version. You don't actually need cells at all. I presume that "cells" in an inorganic version would simply be a label placed on a particular level in the hierarchy. Yep. By cells I mean both neurons and astrocytes. Astrocytes are deeply involved in structuring the background 'blank canvas' that neurons 'write' their EM field interference-pattern message on. But yeah... nature has cells at that level in the hierarchy. It's an amazing, beautiful thing. I am constantly in awe of the elegance of it. These then nest at increasing spatiotemporal scales forming coalitions, layers, columns and finally whole tissue. All inorganically. All the same at all scales from an adaptive control perspective. Power-law scalable. Physically and logically. In my case, for the conscious version the hardware includes the field-superposing, active additional feedback in the wave mechanics of the EM field system produced by brain cells at specific points. The fields form an addition/secondary loop modulation that operates orthogonally, outside/through the space occupied by the chip substrate. Why simulate fields with fields, when computers can do exactly the same thing computationally? Bazzinga. This is a claim that is not proved and that I claim is actually wrong. To prove it you need to do the experiment, not assume it is true. If, as I entertain as a working hypothesis, there is EM field content structured at the nano-scale upwards by interaction with the external world itself, independently of the sensory transduction, non-locally, then "computers can do exactly the same thing computationally" is a false claim. You can't compute the fields because they are an INPUT. A measurement. To prove it, what do you do? You must do the science. Build the same field system and then contrast it with a computed version in an appropriate context. If their behaviour diverges or is indistinguishable then you can say you have done the appropriate science and have answered the question. Not before. What this means in my approach is that the route to real AGI is actually by a physics experiment, with results contrasted with a version lacking the fields. That physics experiment looks nothing like a computer or a robot. It doesn't have to verify 'intellect'. It has to verify EM field behaviour does/does not non-locally couple to the external world. It looks more like the experiments that have currently verified non-locality (entanglement) over 100+ km so far. Only this time done with a chip that produces the field system of a neuron used in the way the brain uses it. There will be a verifiable coupling in the field system separable from the neural activity. Indeed you keep the neural activity and peripheral sensing activity identical and look to see if the field system responds to the distant natural world in repeatable ways. You may think I am off in la-la land. But I say assuming the effect away is in la-la land and that it can be resolved experimentally and that the experiment is decades overdue. If anyone says I am wrong then let me prove it one way or the other in a lab and let the real world determine the result. What I will not do is presuppose a result. I do not know the answer and either does anyone else! Note that you do NOT need to sum the products of voltages times the inverse squares of the distances to every other point in the system, only the inverse squares to the NEAREST points, which themselves form a sort of "shield" from more distant points, and which themselves already contain the effects of the more distant points. However, I wonder what the computationally OPTIMAL thing to do might be, e.g. limit the radius of consideration, etc? You are proposing to simulate 3-D fields in 2-D, but maybe 3-D fields are used only because that is what can be done in wetware. Here, a better grasp of the math involved seems to be in order. The early field implementations will be more 2D/planar. Layering makes it more a 3D field system. Yes, in simulating the way the fields are produced by the membrane, the restriction to a cutoff radius is a valuable way to improve computing loads. Once again: The final device is not simulating anything. Control systems are well understood. However, no one seems to have worked (much) on adaptive control systems theory. What I am starting with is the 'zombie' or symbolically ungrounded version. It doesn't produce the active field system (missing a whole control system feedback mechanism) and uses supervised learning (externalised by a conscious human trainer) to compensate for the loss of the natural role consciousness has as an endogenous supervisor. This sounds considerably MORE difficult than simply putting a 'droid into a real physical environment. Ultimately I am 'putting the droid in a real physical environment'. With a complete kit of I/O. What's missing is the EM field system. Yes the trainer/puppeteer is a difficulty. But it's something we can engineer. This aspect probably needs more elaboration another day (maybe later after the zombie hardware is built). It will, in the zombie form, underperform in precisely the way all computer AGI underperforms. This is what is missing when you use computers to do it all. You end up with a recipe (software) for pulling Pinocchio's strings. Whereas my system bypasses the puppetry altogether. It makes the little boy, not the puppet. Yes. However you view it, there's nothing else there in a brain except nested loops that have power-law responses in two orthogonal axes: sensory and cognitive. My failure/epiphany is that I don't see how cognitive is anything but higher-level computation that considers the prospective effects of potential actions taken to control the system. Remember, "consciousness" probably bares NO resemblance to what is actually happening in our brains that appears to up to have the characteristics we refer to as consciousness. We have dysfunctional models of consciousness that should NOT be carried over into designing future conscious systems. You have a perspective on consciousness that I am having trouble fathoming. I suspect the reverse is also the case! All the evidence so far points to the brain being 'like something' from a 1st person perspective of 'being' the brain. Consciousness. All I am hypothesising that EM fields of the brain (electromagnetism) are what consciousness looks like to a conscious scientist using consciousness to observe it. Why it should be 'like something' from a 1st person perspective of being the fields is our problem as scientists. It's a unique problem in science and its a problem with science, not with nature. This is what my book is about. OK. Let's say your failure/epiphany is X. My logical equivalent failure/epiphany is Y. You can't see how X {etc etc} cannot be the case in the brain.I cannot see how anyone can miss the obvious presence/role of Y in the brain. So there we sit. I can't help this. And I don't know what to do about it except an experimental verification Y is the case or not. All I ask of this forum is that my hypothesis Y be treated the same as X. Not dismissed out of hand for reasons of culture. I ask that the idea of an experiment applied to Y be given the same respect as the decades of inconclusive experiments on X. I am still OK with Y being a failure! What I am not OK with is Y being sidelined by preference of X, not by science. Adding the field system to the sensory axis (e.g. visual experience) or part of the cognitive axis (e.g. emotional experience) provide the active role for consciousness implemented through the causal impact of the Lorentz force within the hardware. You are either seeing something here that I haven't yet grasped, or seeing something here that I have long ago rejected.. I suppose it'd be an 'adaptive control loop' philosophy for cognition and 'EM field theory of consciousness' combined. No computing needed whatever. This is ALL simply 3-D analog computing implemented in wetware. You can name it a computation if you like. You can even abstract it and write it down. That naming/abstraction does not entitle anyone to assume that what is described is not essential physics. That is the notion I have rejected long ago. I found what looks like essential physics. Proving it means building it, not computing models of it. Note that when I build an artificial version of the original natural physics, that too will look like a regularity that could be classified as 'computation' of an apparently identical kind to the original nature. That classification does prove the actual physics any less essential. You have to physically experiment to prove it can be replaced by computation. Remember: I am hypothesising that the EM fields are as essential to a potentially human level robot brain as arm and leg physics are to its body. Just like the brain. Most of the last ten years has been spent figuring out the EM field bits! What have you figured out? How the field is produced by aggregates of ion channels in plaque form. Both the electric and the magnetic field. How it superposes from multiple sources to causally effect neuron behaviour. I attach a video of the electric field system produced by an artificially (randomised) set of ion channels arranged in the membrane (>19,000 of them a rat CA1 pyramidal cell). It operates like a lighthouse centred on the smoa. Sweeping in a loop like a searchlight over its neighbours (that also do it, all superposing on each other). The total field is an emergent property of thousands of little transmembrane dipoles that dynamically coordinate to create a single field system attributable to the whole cell. That field is produced by a single action potential and does not include any contribution from synapses. It is proof-of-principle only. The physics I used was that of inorganic conduction. I am currently setting up to produce the same kind of transmembrane dipoles using convection (electrolyte ions). The result is the same (I expect) but its expression depends on a lot more details like ion concentrations, diffusivity, mobility and the like. Nernst Planck equations. When I have shown how the convection-based fields originate, including the enormous transmembrane electric field, the whole argument will be better founded and I will expect less resistance. That should be done this year along with the beginnings of the zombie chip without the fields. I have done COMSOL simulation that shows, at macroscopic scales, the same dipole field produced by a capacitor that discharges at a highly localised spot in its dielectric. COMSOL turned out to be incapable of also producing the magnetic field (which operates in the plane of the membrane, circulating around the transmembrane current). So I have learned a lot. That I am now omitting, knowing what I lose when I do that (i.e. consciousness). I don't (yet) see the connection between between EM fields and consciousness. If this is truly a hierarchical system, then either EM fields are needed for everything, or for nothing. The EM fields ARE consciousness! You have to BE them, like we do. We are the fields. To be us is to be the fields. There's nothing else there!. It's just that the only fields that add up to anything complex and large and causally efficacious and informative are in the brain. EM fields are a description, by a conscious entity (us), of what consciousness looks like to a consciousness made of EM fields. Think of it this way: there is literally nothing there but electromagnetism. Holding EM fields accountable for consciousness is a no-brainer (so to speak). A choice from a list of 1 is easy. What is not easy is seeing how the fields arise. You can have action potentials without a significant aggregate EM field depending on ion channel co-location and density. You can have action potential and synaptic activity-based significant EM fields that have no causal impact on other cells, depending on cell morphology and cell spacing. Higher cell density means fields have more causal impact. It all requires energy and nature doesn't waste that. So I think it not 'all or nothing' for the brain's particular EM fields. You can have a completely different physical basis for action potential signalling (inorganic, say based on specialised capacitors) that properly configured, produces the same kind of aggregate field system as tissue with the same functional role. The field produced by the ion channels is designed (by nature) to dominate all the chemistry EM field 'noise'. You can make the same dominant field system with organics or inorganics. The total field system is a single unified emergent entity with a life of its own, independent of the underlying neurons. An infinity of different neurons can produce the same field system. A single neuron can produce an infinity of different field systems. Its a vast new axis of tissue operation that I have only barely touched with my work. Teeny weeny Zombie version 0.0 this year I hope. No EM field generation. I call it the 'circular causality controller'. I aim to add the EM fields later. That part requires $millions. ... or as in my recent patent, a clever algorithm that can be efficiently implemented in available hardware. Perhaps one day your zombie chip and mine can be compared (architecture-wise)? Later this year maybe. Mine will be studied/designed using software and maybe the software version will be useful. Don't know yet. I suspect it will never operate in real time. The only thing I know for sure is that my architecture will not function at all without I/O and the real world being attached to it. What I hope to do is commercialise some version of the zombie chip to fund the future non-zombie version (with the fields). I may yet live long enough for that to happen! It's chip-foundry stuff. ANYTHING you can make on a chip can be simulated in a computer, albeit VERY slowly. No foundry would even think of building anything in silicon that had NOT first been simulated. Yes! Likewise no airplane manufacturer would build a plane without simulations to validate the design. But that design is not flight. Making artificial brain tissue is the same. Not being able to simulate the actual field system produced by the chip during operation is not a failure to design the chip ..... any more than the failure of the flight simulator to actually fly is a failure of a design of a flying thing. So chalk me in under this 'adaptive control loop' category for AGI implementation please. I know this forum is a 'using computers to do AGI' forum so I'll just continue to zip it. If these guys are ever to build anything that actually does what they are hoping for, then they simply MUST connect with the computational processes needed to implement such things. If someone here seeks to do something ELSE that is unable to do what adaptive control systems can do, then obviously, it can never ever do what they are hoping for. So, unless Ben, et a., thinks it is out of order here, I think we should look deeper into adaptive control implementations. HOWEVER, remember that what you are talking about IS computation, it CAN be simulated on a digital computer, and the computations can probably be done MUCH more efficiently by divorcing the implementation from the physicality of wetware, e.g. EM fields, etc. The challenge comes in finding mathematical expression of the task at hand, while would (hopefully) lead to a (more) optimal solution (than in wetware). I wonder... If as some here have suggested our computational "goal" is to "understand" things well enough to reduce the information content of what we see to as little as possible, then what better thing to sense than the "noise" in the EM field?!!! If every neuron successfully did this, they would function in the channel-reducing form hypothesized on various past postings. Perhaps the thing that has so limited retrograde propagation NN implementations is that they have lacked this "world view" of their impact on the entire system?!!! The information content in the field system, independent of everything else, is vast by my estimation. When I look at memory and CPU power estimates needed to do real AGI I see it many orders of magnitude underestimated. Take the Kurzweil estimates and put 10 orders of magnitude on it. And it's right there in front of everybody, screaming its presence at us. It doesn't make AGI impossible! It just means you don't base it entirely on computers. Just make the fields the same way... as an adaptive control system ..... voila. I haven't mentioned it much over the years because it seems that most of you aren't interested in my approach. There are many talkers but few doers here. The doers WILL carefully read our postings and accept what makes sense (to them). Our challenge is to "translate" our POV enough so that they can grok what we say from their POV. For reference and for the record.... I am the 'AGI as adaptive control' guy. Correction: Change "the" above to "an". Steve A small and select group. So I guess your adaptive control loops are in software and mine in hardware for reasons of later addition of the field system. So we have that kind of common ground. Am very pleased to not be quite so alone, even tho there are a bunch of conceptual disconnects operating at multiple levels that might separate us. Maybe in time that gulf may be better understood and speak to the wider community in helpful ways. We'll see! I'll report in as results happen. cheers Colin cheerscolin I suspect that when the answers are known, it will be a bit like spread spectrum communications, where there is a payoff for complexity, but where ultimately there is a substitute for designed-in complexity, e.g. like the pseudo-random operation of spread spectrum systems. Genetics seems to prefer designed-in complexity (like our brains) but there is NO need for computers to have such limitations. Whatever path you take, you must "see a path" to have ANY chance of succeeding. You must have a POV that helps you to "cut the crap" in pursuit of your goal. Others here are working on whole-cloth approaches, yet bristle when challenged for lacking a guiding POV. I see some hope in adaptive control math. Perhaps you see something else, but it MUST have an associated guiding POV for you to have any hope of succeeding - more than a simple list of what it does NOT have. Steve AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription ------------------------------------------- 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
