Wednesday, 26 September 2012

The "percepto-bit", can we run our experience in silicon at mesoscopic scales ?

The 2 posts on brain simulation are now rolled into one. Entire Blog on one page at

My whacky idea of the week is a concept called the "percepto-bit" for brain simulation / emulation.  That we can base our box models for consciousness at the mesocopic scale where there is the emergent level of the entire variety of brain functions. Neuromdulation nuclei which can alter emotions to a noticable degree start at mesoscale. Recently I was given a brain tissue lesion counting study for a university project. It turns out it is also mesoscale which is the point we actually start to perceive any problem due to the disappearance of brain areas i.e. lesions, plaques, tangles or vascular blocks when they interrupt memory, emotion and phenomenal consciousness are tiny holes in tissue at mesoscale, but we are not impaired (as far as i know) by these problems when they are microscale level tears in brain tissue. What is modeled below "perceptobit" scale will have to contain the full richness of data flow in a real mind, but since this is below perceptual access substrate, the processing and algorithms can be whatever internal approximations we choose. Only above percepto-bit scale has to be brain accurate. As long as the percepto-bits communicate the known internal processes within these bits to each other, why should we perceive any difference to our own substrate ?

I am not currently sure how original this idea is but it would seem controversial to the low micron level proposed as necessary for simulations/emulation. Last month I decided to just rattle of  a document on what i thought were the main grounds to cover in brain simulation/emulation. Dr Randal Koene an expert in brain emulation then uploaded a video of private discussion to experts in this field, and all the same points had been covered.   

Is mesoscale sufficient to build the substrate of the mind from ? If anybody is familiar with my interest in astroglia field models for columns and Ca2+ waves in neurogenesis, they may be aware, that their properties come together at the mesoscale. Some biophysicist argue that this existance of quantum fields means simulations should be more extreme. i.e. model the brain at nanoscale. If the functional properties are coherent at mesoscale, thats not the case. They actually help bolster the concept to model at larger than micron scale. Perceptobit level is the mesoscale for which any modelling above that is not brain accurate interferes with perception. This is the ground resolution from which information is passed and has to accurately represent brain like signals and physics between each perceptobit. Simulation wise its still pretty intensive to be so physically accurate even at mesoscale, but far easier than microscale.

This had me consider the application of an area I like to specialize in. That the top down biophysics of neural structure might add another simplifying perspective.  I don't propose here not to model low micron scale, just that if we cannot perceive a knockout to anything at that level, then whats in that scale only has to input and output the correct kind of perceptual information to other "perceptobits". So we can just use traditional or new sustrate independent modelling methods within that bit. It is a very strange idea that we could exist and enjoy experience unscathed running in silicon modeled to brain accuracy at sub millimeter fine tip of a pinhead scale. A Very strange idea, even to me, but then so were the previous concepts i came up with and look what happened there ! This reasoning is further explained here at Randal Koenes carboncopies group.  

Can mesoscale simplify proposed nano-scale field effects for information processing ?

Common objections raised to brain sims regard ephatic EM and magnetic fields holding nano-scale information.  This is something I focus on pretty heavily if anybody is familiar with my biophysical proposals for brains structure. If we look at the papers for magnetic mechanisms (folder here), the proposed fields produced by Ca2+ in astrocytes emerge at the mesoscale and are proposed to structure the column itself, while also acting as switching capacitors to sustain sensory signals without them having to constantly re-spike. If we look at some more recent papers for ephatic fields in hippocampus (folder here), and try to build a scenario they appear to be converging towards organizing at theta oscillation in stem cells which eavesdrop for non synaptic go signals.  i.e. Again, these do not appear to be fine grained information fields. Their function like the Ca2+ in the columns may just be to provide a generic mesoscopic coherence. What would be its function ? A linear sort across the entire septal temporal hippocampus axis in sleep is proposed by neuro-computationalists. So the idea is a field activation of Ca2+ flow can restart all the stem cells on that axis at the same time.  Such coherence has a macroscopic function. To re-enstate a line of fresh neurons, facilitates a linear all at once network re-sort along this entire association zone of the brain. If coherence dropped there would be fragmentation of the autobiographical sequences (composed of episodic codings) which forms our sense of self.

The idea is in sub perceptobit level we cannot perceive if the processing is brain like or not. We can implement each bit distributed on hardware in a grid. Even at Mesoscale each bit will require a lot of processing, perhaps an entire cluster by today's standards. There are so many signal types to be accounted for, Glia, modulators, Neurons, enzyme effects, field effects etc.  However one sub perceptual architecture means the requirement of each perceptobit is to present brain like information to its facing neighbor. i.e. Achieve perception by mapping the brain like signals dynamically to each other perceptobit. It can be achived by non brain like computation that fits with the current evolution of computer architecture.

This is speculation of course. I did derive these concepts for field function however by tracking (hypothetically) what I propose are the role of Ca2+ waves in the developmental origins from the radial glia forward in time to when they fade to  become the cortical astroctyes and adult ventricular stem cells. Both my proposals for development and those of the independent researchers adult models citied above are "fairly" consistent with each other.   There is however a recent sub micron electrostatic field model for CaMKII coding the neurons internal cytoskeleton structure by craddock 2012.  If its correct, and the above fields are integrating with this, that maybe problematic. However the end result of that is just the macroscopic neurons structure, so that should be dealt with by sub "pecepto-bit" substrate independent modeling.  A more pressing mesoscopic issue to be considered is that of Glia holding the key to large scale structural plasticity. 

More reason to up the sim resolution. Glia reframe the computational structure of the neuron model.

What does this concept offer if the idea has some truth in it ? What it cannot do is solve the hardest mind upload problem of data acquisition. A top down view, may help in the second major problem of mind upload/simulation. That is the problem of hardware requirements for real time computation in brain emulation/sim can be brought forward by about ten years if mesoscale is enough to run consciousness. Freeing up computational resources can then deal with many simulation /emulation problems like glia and all the unknowns in neuroscience that are vastly underestimated in this field. i.e. Modeling neurons as the calculation factor for a roadmap is not realistic. The current computational evidence shows us neuron graph models structurally are flatpack without glia.

See the neuron graph models with no structure guidance (A) then with structure guidance (B,C,D).   Glia are responsible for self organizing a guidance for the weird morphological structure in the brain, not neurons.  You can get basic signals out spiking neurons networks, but no actual computational brain structures.  

With glia, neurons are constrained by overarching top down principles imposed on them to develop interesting structure. i.e. They can start to build the dynamic, macroscopic complex computational structures we are familiar with. The latest big research from brain transcriptome centers are unable to determine known cortex structures from the genetic mapping of neurons. What does this mean ? The neuron as the basic component for simulation/emulation is to a degree a faulty concept, but the degree to which it is faulty has amazingly not been realized and our largest simulation/emulation projects are "trying" to run on it ! We do get oscillations, and crude spike co-coalitions, but anything else remotely brain like has to be arranged for the network by programming structure. It does not derive a large brain like network structure from internal basic principles. Why ? A great problem is the various glia have their own unrecognized structure building principles (the focus of my PhD research) which is forcing the greedy growth principles of neurons into large complex structures that are macroscopic scale. So there is then no single entry point or morphology to pick for brain simulation. Ideally then model all the morphology at all levels but being practical thats too power hungry. So a proposal is to realize the problems with the neuron substrate model and compromise to a box substrate model comprised of the baseline size for perceptual bits which communicate the known internal processes (including glia, Ca2+ waves ,+ room for unknowns etc) within these bits to each other. 

This mesoscopic scale may not be sufficient, or there may have to be variable scale focus for different brain regions, but if the top down concepts can simplify this problem and bring simulation/emulation timelines forward, why not put the idea out there for consideration.  I don't even know if this is an original concept. The idea does seem pretty obvious.
Some introductory  articles on glia. For more on ramifications of glia see this site. Much of the theory on this site is also about the structure that glia builds.

Without Glial Cells, Animals Lose Their Senses

This article shows how neurons collapse without glia, pretty consistent with the computational graph models

Glia Guide Brain Development In Worms

Even in worms the Glia are guiding structure. It is the radial and various glia which gives larger brains their structure.

Brain's Connective Cells Are Much More Than Glue: Glia Cells Also Regulate Learning and Memory

No surprise there. The Ca2+ astrocyte wave models are proposed as a key to short term memory, as well as hippocampal neurogenesis.

Thursday, 16 August 2012

New paper Modality Independent Neuro Development Substrate for Artificial General Intelligence “MINDS for AGi”

Here is my application of the dipole / multipole framework to AGi for AGI-12, 5th Conference on AGI @ Oxford, Dec 8-11 2012  -- POSTER LINK

Its called The Integration of a Deep Structure Neuromorphic Framework for AGi: Modality Independent Neuro Developmental Substrate for Artificial General Intelligence “MINDS for AGi” - LINK

2b edited 30 pages! 

Whats it about  ? If you have the entire structure principles, you have generalized map for the entire computational principles. When filled out with low level detail its perfect for the highest level modality free neuro models for AGi. Needs refinement to specific in depth computation, but a start towards high level principles. 

Also brief 2 page outline for application of Dipole expansion framework to brain emulation / copies - LINK

Felix Lanzalaco (in picture) presents his long awaited abstract to a paid audience outside in the cold scotland climate.  Some youtube levity was needed after writing such a heavy going paper, but is she on to something ? Even more to the point, am I on to something, OR did I slip into groupthink and/or pseudoscience (yes the video was improvised).


UPDATE 16/09/2012 : As predicted by all the paper was way too long for AGI-12 as its a Springer proceedings Journal with fixed length.. The content itself passed review on all 4 measures used (That surprised me !), but both reviewers rejected on the editorial issue of length. It has not been possible for me to reduce it to 10 pages right now for these reasons (aside from my final neuroscience exams for this year also).

1. Since producing it I have realized the concept still needs fleshed out further, i.e. Its still an over-simplification too far. 

2. To shorten to a core would need reference to its original larger work, and this has not been submitted to a repository.

To my surprise one of the reviewers proposes high level architecture is harder to get right than sub system components. Thats what I advocate also, but other AGi programmers think differently on their quest for simple principles. They haven't experienced enough neuroscience in my view. This review point is the primary reason I could not compress the conceptual framework right now. The unique insight I propose to posses into the highest level structure is not enough. There has to be more work to produce consistency with both known principles and known unknowns in neuroscience at all key levels of scale before a compression to high level principles. I already had the plan, and carried out groundwork in how to go about this, thats what the PhD is about. Its no easy task and requires a thesis just to cover the projects scope. 

Sorry to say but my current opinion is if the brain is anything to go by even with a full understanding at all levels the final principles for general intelligence are going to have to include the integration of a lot of sub principles, and that would be after such principles have been well untangled from the biological substrate. Maybe there is an E=mc2 for the brain, but I have doubts it can capture the true systems functionality. My hope is the neuro-computation aspect of the biophysical understanding I bring will be similar to the resultant success I experienced with the developmental predictions made almost 10 years ago (from the proposed top down models here).  That is that having the final high level insight integrating with the lower-level work of todays neuroscience can make predictions which can inform us in our AGi (brain derived) system research regarding both known/unknowns and the hardest of all, unknown/unknowns.

Thursday, 31 May 2012

Visual morphology used as method by Izhikevich's cutting edge "Brain Corporation" labs

Dr Eugene M. Izhikevich, the creator of scholarpedia, runs one the worlds leading projects for transcription and simulation of the Mammalian brain in a computer system. He recently employed Dr Hermman cuntz who provided a "greedy growth" algorithmn which describes the structure of neurons based on how they grow with any given space (paper here).

He clearly states "Morphology is key to understanding both circuits and computation, since it reflects the constraints given by both"

or from the video at end of this post

“to have a good model, you need to have information about the shape of the cell (neuron)”.

Below is the result of using this structuralist principle by placing function derived from structure algorithms in cell spaces such as the hippocampus, cerebellum etc.

Paradoxically Nobel Laurate Gerald Edleman who assisted Dr Izhikevich in 2009 on these projects, like many others before and after him, have told me to ditch the approach of using morphology as a guide to function, when i presented the dipole cortex concept to the journal he was working on at that time. Perhaps all members of the team do not see eye to eye on all matters "brain".  Regardless of disagreements between new and old schools of thought, the fact is that the work of Izhikevich, Cuntz and Brain corporation relies heavily on structuralism.   By doing so they have with a fraction of the budget, produced results that exceed the massively funded human brain project which follows a completely gene-centric position.

The following video about brain corporations work is well worth watching. I have been criticized repeatedly for asking questions of does the cortex dipole like morphology dictate its computational function. Here we see this approach being advocated just now for neurons. I hope to try and persuade these groups at some point we can use Magnetohydrodynamic (MHD) principles within similar constraints to entire structures such as the cortex, cerebellum, limbic regions etc in neurodevelopment, and that the MHD roots of neurodevelopment also has the potential to reveal the still unknown principles behind the parallel wiring of axon bundles.

i.e. to understand connectivity, Izhikevich teams up with Frank C. Hoppensteadt to explore the abstract concept of poly-synchronous wavefronts (here). This is compatible with my developmental view that brain morphology that has highly dense axon connection's with continuous synchrony (such as the limbic areas) are dominated by harmonic interactions that have their roots as Magnetohydrodynamic (MHD) harmonic modes which arise in the ventricular zone. This image gives a visually simple perspective of the concept of such modes.

For more in depth information, I have updated my 2009 post where i show differences between the MHD work I have advocated, and independent findings (based on traditional turing model for development) which have evolved laterly to verify at least part of this concept.

NOTE: I have also updated the findings after i presented my work at the conference "integrative approaches to brain complexity".  i.e. Its become clear to me now why these brain simulations projects will fail without top down models.