Connectionists: How the brain works

Fri May 23 14:16:49 EDT 2014

I offer some thoughts on this, extracted from a recent paper:

“In fact, we find that actual CNS neurons are endogenously active, with baseline rates of oscillation, and with multiple modulatory relationships across a wide range of temporal and spatial scales:
silent neurons that rarely or never fire, but that do carry slow potential waves (Bullock, 1981; Fuxe & Agnati, 1991; Haag & Borst, 1998; Roberts & Bush, 1981);
volume transmitters, released into intercellular regions and diffused throughout populations of neurons rather than being constrained to a synaptic cleft (Agnati, Bjelke & Fuxe, 1992; Agnati, Fuxe, Nicholson & Syková, 2000); such neuromodulators can reconfigure the functional properties of  “circuits” and even reconfigure functional connectivity (Marder & Thirumalai, 2002; Marder, 2012);
gaseous transmitter substances, such as NO, that diffuse without constraint from synapses and cell walls (e.g., Brann, Ganapathy, Lamar & Mahesh, 1997);
gap junctions, that function extremely fast and without any transmitter substance (Dowling, 1992; Hall, 1992; Nauta & Feirtag, 1986);
neurons, and neural circuits, that have resonance frequencies, and, thus, can selectively respond to modulatory influences with the “right” carrier frequencies (Izhikevich, 2001, 2002, 2007);
astrocytes that[1]:
·      have neurotransmitter receptors,
·      secrete neurotransmitters,
·      modulate synaptogenesis,
·      modulate synapses with respect to the degree to which they function as volume transmission synapses,
·      create enclosed “bubbles” within which they control the local environment in which neurons interact with each other,
·      carry calcium waves across populations of astrocytes via gap junctions.
These aspects of CNS processes make little sense in standard neural information processing models.  In these, the central nervous system is considered to consist of passive threshold switch or input transforming neurons functioning in complex micro- and macro-circuits.  Enough is known about alternative functional processes in the CNS, however, to know that this cannot be correct.  The multifarious tool box of short through long temporal scale forms of modulation — many realized in ways that contradict orthodoxy concerning standard integrate and fire models of neurons communicating via classical synapses — is at best a wildly extravagant and unnecessary range of evolutionary implementations of simple circuits of neural threshold switches.  This range, however, is precisely what is to be expected in a functional architecture composed of multiple scale modulatory influences among oscillatory processes (Bickhard, in preparation-a; Bickhard & Campbell, 1996; Bickhard & Terveen, 1995).

[1] The literature on astrocytes has expanded dramatically in recent years: e.g., Bushong, Martone, Ellisman, 2004; Chvátal & Syková, 2000; Hertz & Zielker, 2004; Nedergaard, Ransom & Goldman, 2003; Newman, 2003; Perea & Araque, 2007; Ransom, Behar & Nedergaard, 2003; Slezak & Pfreiger, 2003; Verkhratsky & Butt, 2007; Viggiano, Ibrahim, & Celio, 2000.“

This is from:
Bickhard, M. H.  Toward a Model of Functional Brain Processes: Central Nervous System Functional Architecture.

The basic critique of information processing models applies just as strongly to Predictive Brain models.  It is elaborated, and an alternative is outlined, in:
Bickhard, M. H.  (forthcoming, 2014).  The Anticipatory Brain: Two Approaches.  In V. C. Müller (Ed.)  Fundamental Issues of Artificial Intelligence. Berlin: Springer (Synthese Library).

I have attempted some first steps toward both a micro-functional and a macro-functional model that can account for and accommodate such facts in the functional brain processes paper.

Mark

Mark H. Bickhard
Lehigh University
17 Memorial Drive East
Bethlehem, PA 18015
mark at bickhard.name
http://bickhard.ws/

On May 23, 2014, at 12:45 PM, Brian J Mingus <brian.mingus at colorado.edu> wrote:

Is there anything that can't be represented as a single equation or a really long run on sentence aka model?

With regards to whether new models such as *poesis are accepted by the field, I think this really boils down to identity politics. Most researchers are doing a mix of wanting to understand how their brain works and wanting to help humanity by grokking the brain to solve problems like epilepsy etc. The part of them that just wants to understand how their brain works will obviously tend to prefer their own personal rotation of the space. The part that wants to help humanity sees the need to integrate new theories, but this conflicts with the ego and these new theories are more likely to be changed beyond recognition to fit into a given researchers existing framework than to be supported and encouraged.

Every once in a while a researcher will stumble upon a description so short and so elegant that it easily transcends the usefulness of all existing theories. However, the brain isn't like previous objects of study. It's essentially the most sophisticated thing that science has ever turned its gaze on. Whether it lends itself to simple "single equation" descriptions is an open question, but I personally doubt it. All models are wrong, and this applies to every model of the brain. Some models are useful, and this also applies to every model of the brain. For the purposes of understanding how your own brain works, an arbitrary rotation of a sophisticated theory seems quite sufficient. For solving actual problems, like epilepsy, some models will be more useful than others. That said, a model isn't always even needed to solve a problem: the latest epilepsy drugs are the most effective and they are found by shotgun approaches which result in drugs that work and for reasons that nobody understands.

Zooming out, I like to ask myself whether there's a reason things are the way they are. This is obviously an unanswerable question, but it does shine a light on the fact that we have egos, and that this process of ego-scaffolding leads to many researchers focusing on different perspectives at different levels of analysis. The academic publishing system then broadcasts these perspectives, and whether or not we give truly fair credit assignment versus implicitly mashing their theory into our own preferred framework, everything does in the end get all mixed up together, resulting in a better set of theories overall. One wonders if this apparently fortuitous coincidence isn't a coincidence after all. I personally suspect the usefulness of what we're doing is that we are all contributing to the building of something great. While this somewhat justifies the existing system, given that it actually appears to be working (in a way that none of us understands), I would also advocate a more egalitarian approach where we open our minds to as many theories as possible and cheer on perspectives different from our own. And from this angle, I really like the promise of more informal mailing list conversations for the spread of ideas and hope that you guys keep it up, because I love reading your ideas. And I wish more folks would jump in too and help us all out, rather than just hiding out in the darkness of the literature!

Brian

https://www.linkedin.com/profile/view?id=74878589

On Fri, May 23, 2014 at 8:43 AM, Hans du Buf <dubuf at ualg.pt> wrote:
I was out for some time (mailbox overflooded) and now saw again discussions
about vision and motor control and single equations and and and...
Why don't you start with the archaic part of our brain? (NOT the frontal lobe
and rich club - the massive communication hubs; white matter - only these
distinguish us and great apes from rodents)
I've been reading a few recent reviews, and the idea comes up that the
enormous complexity is the astonishing result of merely replicating very
few structures over and over (single equation :-)
We know the laminar structure of the neocortex and the connections and
processing (FF input from a lower level, horizontal processing, FB input from
a higher level), the hierarchies V1 V2 etc and M1 M2 etc and A1 A2 etc.
Oops, FF = feedforward.
These hierarchies are reciprocally connected: V2 groups features from V1,
V4 groups features from V2, until IT cortex with population coding
of (parts of) meaningful objects, but in each step up with less localisation
(IT knows what the handle and shank of screwdriver are, and about where,
and that they belong together; but don't ask IT to put the tip into the slot
in the head of a screw - for that you need V1, but V1 has absolutely no clue
what a screwdriver is). FF+FB is likely predictive coding with a generative
grouping model. If you have V1 and V2, you also have V4 and IT. You can
also assume that the processing in the V and M and A hierarchies is the
same: one equation.
All neocortical areas are reciprocally connected to pulvinar (LGN in case of
vision) and higher-order thalamic areas in a laminar way, and then to basal
ganglia via layers for arms, face and legs. The BG take decisions, most important
keyword: DISinhibition. One equation.
All visual areas are still connected to motor areas (archaic brain, rodents,
screwdriver).
All motor areas are connected to sensory areas: corollary discharge signals
were first introduced because of saccadic eye movements, but they are
ubiquitous for distinguishing external from self-induced percepts, and at
all levels: from reflex inhibition, sensory filtration, stability analysis up to
sensorimotor learning and planning. I boldly assume: one equation.

Once you understand this, you could assume the same principles for other
cortices, like the anterior and posterior cingulate: ACC for arousal and
attention, error, conflict, reward, learning; PCC for more internal attention
and salience. The PCC is connected to thalamus and striatum (basal ganglia,
decisions!). ACC+PCC balance internal and external attention, both between
narrow and broad attention. Internal: daydreaming, freewheeling, autism.
Fronto-parietal network: short-term flexible allocation of selective attention.
Cingulo-opercular network: longer-term, maintain task-related goals.
They interact via cerebellar and thalamic nodes, work in parallel.
They are part of the default mode network: external goal-oriented action
vs. self-regulation. Homeostasis. Small imbalance between endogenous and
exogenous processes: ADHD.
Anterior insula and dorso-lateral PCC: balance between excessive control and
lack of control. Imbalance: obsessive-compulsive disorder or schizophrenia.

Keep in mind that our brain is always testing hypotheses and predicting errors,
always at the brink of failure. Metastability: shift through multiple, short-lived
yet stable states. You tweak a parameter and the brain freaks out. It is amazing
that (in my view) a very few principles could be applied to understand how our
brain works - and that most brains seem to work quite  well.

Finally (I need to get some work done), the brain is not a bunch of artificial neural
networks which are trained once. It constantly re-trains itself like a babbling baby,
although this is rarely noticed.
Am I too bold?
Hans

On 05/23/2014 03:27 AM, Janet Wiles wrote:
Will a single equation be a good model of the brain as a whole? Unlikely!
Will a set of equation-sized-chunks of knowledge suffice? It works for physics, but remains an empirical question for neuroscience.

The point is not about what's the best way to model the brain, but rather, what models are adopted widely, while others remain the province of a single lab.  A model that can be expressed as a single equation seems to be a particularly effective meme for computational researchers.

Janet

=======================================================================
Prof.dr.ir. J.M.H. du Buf                          mailto:dubuf at ualg.pt
Dept. of Electronics and Computer Science - FCT,
University of Algarve,                            fax (+351) 289 818560
Campus de Gambelas, 8000 Faro, Portugal. tel (+351) 289 800900 ext 7761
=======================================================================
UALG Vision Laboratory:            http://w3.ualg.pt/~dubuf/vision.html
=======================================================================

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20140523/50e18df4/attachment.html>