It would be fair to say that I am absolutely enthralled by the workings of the brain.
For me, the brain represents the epitome of complexity. Everything that we humans are, everything we do, all of our feelings, our thoughts (thought itself), our behaviors, all are due to the workings of that marvelous roughly kilogram and a half of gelatinous living mass. Human behavior is infinitely complex in details. Unpredictable, often. Yet there are universal aspects to behavior that should resolve down to how the brain works.
If we have any hope whatsoever of understanding ourselves as a species, and appreciating our strengths and weaknesses as a collective society as well as individuals, we need to understand what is going on in our brains. Progress is being made. Whether it will produce adequate results in time to formulate plans and policies regarding how to mitigate our species' weaknesses before we wipe ourselves out is another question. Nevertheless, our neurosciences are delving deeply into the workings of the brain and we are learning. What stands out the most, right now, is the sheer complexity of how a bunch of living cells can, ultimately, produce human behavior.
For most of human history, we have largely assumed that our mental strengths, namely problem solving intelligence and creativity, can and have overcome any weaknesses. The latter have been relegated to our emotional side. We have assumed that our ‘darker side’ is a left-over remnant of our animal evolution (or if you are of a religious bent, perhaps it is a result of the free will thing). The popular (both in philosophical and vernacular realms) view that human intelligence can subdue our animal spirits and this is what leads to civilizations based on the rule of law has dominated all of our thinking about what we humans are, how to explain our history, and what our destiny will be, especially if we work hard with our rational minds to overcome our emotional minds. The creation of the fictional character Spock, on the television series Star Trek is perhaps the best example of this self-image zeitgeist. Spock, from the planet Vulcan, was a being of pure logic (except that he was half human and the ‘lesser’ qualities would sneak out from time to time to make the plot interesting). As such he could be counted on to think through any conundrum and save the day. The idea that humans might, one day, evolve to have this mental capability was, and still is, appealing since we can easily point to all of our failings arising from the negative emotions.
But evolutionary neuroscience and psychology are showing us more and more that this notion of a perfectly logical being is not only not likely, it is not even desirable. Emotions and affective states in general are a necessary function of the brain. They are essential to most forms of decision making for the simple reason that most decisions are made under considerable uncertainty. And emotional responses to situations that are similar to other prior encountered situations that had an emotionally meaningful outcome can be very helpful in making rapid assessments and decisions (see Antonio Damasio's “Descartes' Error” for a full explanation).
Evolution found another way to balance emotional influence over intelligence (decision making) in the form of sapience (see my working paper series on sapience and wisdom, here), the capacity to override or modulate emotional responses and bring tacit knowledge to bear on decisions; what we call judgment. Sapience is the brain basis for gaining wisdom over one's life. As one experiences the vagaries of life one gains tacit knowledge (subconscious knowledge that can be applicable to a large variety of situations). In more mature adults this produces the various behavioral attributes we recognize as wisdom. The capacity for sapience is the newest evolutionary step in making humans the creatures they are. It, like affect, shapes intelligent decision making, but based on knowledge acquired during one's life, not evolutionarily hard coded into the limbic brain. Unfortunately, as the newest brain faculty, it is also the weakest (not as influential as either affect or the logic of intelligent/creative problem solving.) Presumably our social organization with its dependence on cooperation for group survival would eventually lead to stronger capacities in this area. Sapience is a brain function that could undergo further evolutionary strengthening given the chance. Indeed, there is some growing evidence that the there is a reasonably wide range of sapience strength in the population already. Just as there is variation in intelligence and creativity, the same seems to be true for sapience as well. A great deal more work in neuroscience and psychology will be needed to clarify this situation and some of that, particularly in the area of judgment capacity, is already underway.
Meanwhile one can only look at the complexity of the brain and marvel at the profound nature of mind. In this post I want to view that complexity from a hierarchical scale perspective. That is, by examining some of the issues of complexity at various scales of organization we can begin to see how the overall functioning of the brain is the result of complex processes operating at all of these scales and the interactions between them.
Complexity
It must be the height of audaciousness for we humans to actually believe we can understand how the brain works! And yet it often comes as a surprise to many non-neuroscientists just how much we do understand about the brain's workings. Any complete understanding is still far off, but there are glimmers of hope. For one thing we have a reasonably good handle on the scope of complexity involved in brain functioning. By this I mean the way in which the brain is a complex system comprised of complex subsystems, which, in turn are comprised of complex sub-subsystems. The state of neuroscience, viewed from a macro-level, shows a hierarchy of complexity from the molecular functions within synaptic junctions (and associated glial cells) through the workings of whole neurons, through circuits of neuronal assemblies, all the way up to the functioning of whole brains comprised of complexly intercommunicating modular regions.
There are still many details not yet understood, such as exact wiring diagrams for circuits and regions. We still have a long way to go to delineate the intricacies of all of the neuromodulator molecules and their effects on various kinds of neuron types. Even the zoo of neuron types is probably not complete. But nevertheless the broad outlines of some very powerful principles appear to be taking shape that allow us to produce better hypotheses leading to better experiments and yielding more useful information per unit time.
In this somewhat lengthy blog, I would like to provide a descriptive story of the levels of complexity of the brain from what is generally now accepted knowledge in neurobiology. However, interwoven within this story I will have to include my own speculations/observations that might help tie the various thread levels together. The reason I'm jumping into this fray is that I am not a neurobiologist with a reputation to protect, who studies a specific level of phenomenon from a reductionist perspective. I am, on the other hand, a dedicated scholar of all things brain. I have been deeply involved in tracking developments at all levels of brain complexity for many years. And I have no particular bias regarding any particular aspect of brain science save my intense interest in sapience. When I go on a speculative tangent, I will try to flag it as such so you, dear (patient) reader, will be able to judge the story accordingly.
Figure 1, below, shows a schematic diagram of brain organizational complexity so far as it is understood today. There are multiple different dimensions of complexity and here only three are shown. The first dimension of complexity in the figure could be called levels of organization, from the synapse through to functional neuronal clusters such as nuclei (globular-like clusters) and cortices (sheet-like clusters). These levels are dependent on those below as indicated in the concentric ovals. Synaptic complexity is, itself, dependent on chemical processing complexity at a lower level than shown. The second dimension could be called functional identity; various brain regions process different kinds of information even if the underlying mechanisms for doing so are the same as delineated in the first kind of complexity. Finally, the third dimension of complexity is indicated as the inter-module wiring organization or schema that produces the whole brain activity and the animal's behavior. All of this is shown as against another dimension, that of genetic and developmental complexity, which will not be covered in this blog.
Figure 1. A schematic representation of three dimensions of complexity in the functioning brain, against the background of genetic and developmental complexity.
Other dimensions of complexity, to some degree, map onto these dimensions rather naturally. For example the temporal dimension roughly maps onto the levels of complexity with the fastest time scale represented by events in the synaptic junctions (chemical reactions). However, it turns out that longer-term processes also impact the synapses as is the case for the development of long term memory traces (see below). Longer temporal scales are involved in neuronal activities, and longer still scales in circuit activities.
Another important dimension of complexity not addressed in the schematic is genetics and particularly the role of evolution and development on the shaping of the phenotypic form. To some degree these can be seen in the various functional modules, their sizes and relative processing powers being due to both genes (evolution) and development influences (e.g. epigenetic factors).
I will attempt to tie these into the story as they seem fitting. But this story will largely revolve around the diagram as a kind of backbone on which to hang the flesh, as it were.
There have been a spate of predictions from the artificial intelligence crowd over the years about when the computational processing power of computers would overtake the capacity of the brain (c.f. “Mind Children by Hans Moravec). Relying on the so-called Moore's Law phenomenon as an indicator of processing power, and a very overly simplistic estimate of neuronal processing power (in megaflops per second — a wholly unjustified comparison), these futurists estimated that we would soon be able to build computers that exceeded the human brain's abilities. Aside from the fact that they started with almost no real understanding of just what sorts of algorithms they might need to emulate to get that kind of ability, they seriously and grossly underestimated the informational complexity of just what is going on in the brain in the first place. If one were to hazard a guess at the processing capacity of the brain, they should start by looking at the complexity of just a synapse all by itself. Neural network modelers are infamous for trying to simplify the synapse down to a simple scalar variable (the synaptic weight) and a linear learning equation that would change the weight according to simplistic interpretations of Hebb's rule. In fact the complex chemical processing taking place at the synapse is extremely difficult to emulate in our very fastest computers. The best we can do is approximate some functional form of what it is doing to form and maintain memory traces. The real thing is wildly complex. Just ask any neurobiologist.
The Synapse
Synapses are the junctions between neurons. First note that there are actually two major categories of synapses, electrical and chemical. Actually both have chemical and electrical phenomena going on so that distinction is a little confusing. Chemical synapses probably ought to be called semi-discrete or pulsed synapses since they operate on the basis of pulses of electrochemical discharges of neurotransmitters. Electrical synapses might better be called continuous or gradient synapses since they communicate information more like an analog device. The problem with either of these descriptors is that they don't quite own up to what the phenomenon actually entails. For example, in the case of the pulsed synapse, the release of neurotransmitter chemicals into the synaptic gap leads to an analog build up. If the pulses come in rapid succession, there is something like a gradient response in the post-synaptic membrane. Nature, it seems, could not make up her mind in terms of the discrete signal vs. analog and so chose to create two versions of a mixture between the two. It could be that discrete-like pulsed signals are more reliable over long distance communications channels (the axons of the neurons - all of the electrical synapses are between near neighbor cells). But analog (continuously fluctuating) signals are needed for reliable storage and integration of state information.
Figure 2 shows a (still oversimplified) diagram of a typical chemical synapse. An axon or ‘cable’ from another neuron, or a sensory end organ, conducts a pulsed signal called an action potential. This is an electrochemical wave that travels down the axon (always away from the origin) and terminates in a synaptic bud. The latter contains packets of neurotransmitters of various kinds depending on the type of neurons involved. There are many different kinds of neurotransmitters and many more kinds of neuro-active molecules (also called neuromodulators). Some neurotransmitters stimulate the receiving neuron to fire, others inhibit it. Right here we find many sub-dimensions of complexity. The dynamics of post-synaptic membrane behaviors are highly variable and dependent first on the kinds of neurotransmitters and modulators found in the vicinity of the post-synaptic membrane.
Figure 2. The synapse is an incredibly complex, dynamic actor in the whole panoply of neural processing. Here chemical and electrical forces interact to produce cell to cell signaling with signal processing, such a noise filtering, and even noise introduction for some purposes. The synapse is the site of memory trace encoding in that the receptivity of the post-synaptic membrane increases under certain signal and associated conditions. See text for details.
Synapses can form just about anywhere on the receiving neuron's body, but in most instances form where an axon comes in contact with a dendritic spine (as in the figure). On the arrival of an action potential (1) the pre-synaptic boutton (bulge) releases the packet contents into the gap between pre- and post-synaptic membranes, where they rapidly diffuse across and couple with receptor sites on the post-synaptic membrane (2). These receptors are tied to protein channels that open (or close depending on which neurotransmitters) allowing the influx of sodium ions (3) into the interior of the post-synaptic compartment. This sudden influx of positive ions changes the electro-potential across the membrane which can cause various voltage-sensitive channels to open as well (4). There are actually a number of other ions that may cross the membrane in response to a signal, but the important aspect here is that the electro-potential is changed dramatically. This depolarization wave can travel down the dendritic spine toward the cell body (shown as an arrow from #3 to a wave form traveling toward the receiving neuron's main body). When I get to the neuron level of complexity I will further elucidate what happens then. For now it is only important to recognize this and certain chemical changes in the immediate post-synaptic compartment as real-time responses to the incoming action potential signal.
One major real-time effect is for calcium ions (Ca++) to enter the post-synaptic compartment (5). This starts a series or cascade of chemical events that operate on a slower time scale (6) than the real-time events and depend on the accumulation of calcium ions over time. The ‘intensity’ of a pulsed neural signal is encoded in the frequency of action potential arrivals. The more frequently that they arrive at the synapse, the more calcium ions build up (and are slow to be removed) and thus stimulate several different second messenger effects (7 and 8). One effect (7) reinforces the sensitivity of the post-synaptic compartment as a form of short-term trace of the preceding action potential stream called short-term potentiation. This simply means that the synapse becomes more likely to fire a depolarization event even with weaker (less frequent) subsequent inputs until the cell manages to sequester or remove the accumulated calcium. Effectively the calcium buildup acts as a kind of leaky integrator or capacitor storing a short-term memory of recent events (see this Wikipedia article for more details). At the end of a long chain of second messenger events and in correlation with other chemical conditions brought on by either the activities of nearby synapses or by neuromodulators in the extra cellular matrix, chemical signals are sent to the protein construction mechanisms (called ribosomes - 9) and back to the cell body and to the nucleus (10). The former would appear to reinforce instructions to keep producing proteins needed to keep the channel concentration up to snuff. The latter is thought to be activating genes in the nucleus to increase the production of messenger RNA (somehow tagged to be delivered to the synaptic compartments that sent for it!) that will up the concentration of channels in the post-synaptic membrane (11, 12, and 13). These membranes undergo structural changes that are very long lasting and constitute longer-term strengthening of the memory trace. Synapses, so strengthened, are capable to generating huge depolarizations even with weak incoming signals, and even after long periods of quiescence.
Synaptic dynamics and the encoding of memory traces in synaptic strength in several different time domains is the basis for memory phenomena at the circuit level as I will discuss below. But as you can see, already we are dealing with immense complexity (chemical and temporal) and we haven't even considered what is going on in the rest of the neuron. We'll now consider a bit of that.
The Neuron
The first thing to understand is that there are many kinds of neurons in the brain (Figure 3). I can barely do justice to the variety and what is known of their different functions. Here I will just consider a single type of neuron, one that is common in the cortical structures discussed later. That is the pyramidal cell (see upper left corner of the figure below). This kind of neuron, of which there are sub-types depending on what cortical area one is looking at, appears to act as a major integrator of diverse convergent signals from both local and distant neurons.
Figure 3. Various neuron types found in the brain. Image from Consortium on Cognitive Science Instruction (image)
Almost all of these types of neurons derive from a base-type and they all have many features in common insofar as their functioning for receiving and sending signals. Figure 4 shows what we might call a typical arrangement for neurons in their role as communications devices.
Figure 4. Typical neurons receive input signals (action potentials) in the dendrites or on the cell body and send signals down the axon toward other neurons or end effectors (like muscles). A fatty sheath (myelin) insulates the axon much as a plastic coating insulates an electric wire but also helps to speed the signal in those axons that are so equipped. Not all axons have sheaths, however. Generally only long distance axons (from one brain region to another) are sheathed. The bundles of long-distance axons form what is known as the white matter in the brain. Neuron bodies form the grey matter. Image courtesy Tutorbene (image).
Neurons are said to ‘fire’ each time an action potential is sent down the axon. Whether or not a neuron fires depends on a complex integration of all of the incoming signals received in the dendrites and cell body synaptic junctions. At the root of the axon (see slightly darkened region of the cell body above where the axon emerges) a ‘threshold’ function determines if the sum excitation from the inputs should, at a given instance in time, generate an output action potential. If the inputs are weak, or out of synchrony, the neuron can fail to fire or fire very sparsely (low frequency). On the other hand, if the inputs are strong and closely aligned in the temporal domain, the neuron can fire vigorously (high frequency). Thus, the neuron is a relay, a filter, and an amplifier depending on the total input activity. Basically the dendritic processes and synapses provide a temporal integration over all incoming signals. The neuron's dendritic tree structure and its body membrane provide a spatial integration over all incoming signals. This is how it is possible for even a single neuron to act as a pattern recognizer.
The dynamics of neuronal firing is based on all of the previously mentioned complexities of synaptic dynamics plus those resulting from the complexities of the cell itself. Synapses that strengthen due to high levels of activity at critical moments when other inputs have excited a particular cell (or even just a particular branch of a dendrite!) produce a memory trace through the neuron. If those synapses have a tendency to excite the cell body they can produce an output signal in the cell whenever their source cells are excited. It is the case that a single neuron may participate in many different memory traces, acting as a convergence zone for different memory features that constitute a cause for activating whatever other cells the axon runs to.
Neurons participate in networks in which each neuron acts as a little integrating computer of sorts. Many neurons can innervate a single neuron, which, in turn, can activate many other neurons. The dendrites pull in signals from all over, both local (nearby) neurons and distant ones. The axons generally branch out and run to many other neurons, again, both local and distant. The possibilities are literally infinite, especially since we now know that neurons are forever forming new connections even while breaking off old unused ones. The brain is continually reforming as new memory traces are built from new experiences. At the same time it reinforces memory traces that have proved to be useful in living experience. But each kind of neuron is a complex processing unit in itself. We are just beginning to unravel some of the mysteries of various sub-functions that the different types of neurons perform, their dynamics, and their interactions with one another.
Interested readers may find more information about synaptic and neuronal processing in my paper on simulating these, “Toward a Theory of Learning and Representing Causal Inferences in Neural Networks”.
The Local Network
Neurons work together in complex networks to process spatio-temporal patterns of inputs (say from the senses), associate these patterns with context (other patterns) as well as with the internal state of the organism (drives, emotions, ideas, etc.) to produce meaningful output (e.g. behaviors). In the basic sense, meaningful output means producing behaviors that support the fitness of the beast in question. It turns out that many researchers have demonstrated how neural networks and their activities produce meaningful behavior in various animal models. While this is still preliminary there are many really convincing demonstrations of how neurons (including primitive brains), working together in specifically organized networks, produce fit behaviors.
One of the most primitive networks that evolved for the purposes of movement control is called a ‘central pattern generator’ (CPG). This is an arrangement of neurons that mutually excite and inhibit dynamically in such a way as to produce an undulating (not-quite sinusoidal, but cyclical) wave when innervating opposed muscle groups. I, along with my colleague Paul Fisher, explored this phenomenon when developing a search control strategy for our robot MAVRIC . Our published paper, “Foraging Search at the Edge of Chaos” provides a detailed description of simple CPGs as found in nature and the one we simulated (using the simulated neuronal and synaptic dynamics described above) to produce the ‘drunken sailor walk’ search pattern described in that paper. CPGs of various kinds are responsible for most of the kinds of dynamic muscle coordinations needed, for example, to generate varying gaits in running. The same circuit can respond to different input signals by changing average frequency and amplitude of the wave pattern output. This is an example of nature's phenomenal way to provide multiple functionality out of single components.
CPGs generally do not involve a great deal of learning. They are basically multi-phasic oscillators that switch modes depending on variations in input (through synapses from other circuits) signal. Those inputs, however, can be subject to long-term modification due to learning, i.e. memory traces encoded in the synaptic chain in other circuits. The above paper on Causal Inference also explains the basic neuronal unit of networks that learn associations between signals that are hard wired to convey semantic information (e.g. the smell of food) and signals that come from incidentally firing circuits. If the latter consistently fire a short time prior to the firing of the semantic signal, then a longer-term association is encoded between them, such that the incidental signal (called a conditioned stimulus in the psychology literature) may be sufficient, by itself, to cause the receiving neuron to fire, leading eventually to a motor response (the conditioned response).
Another wonderful example of a local circuit network that preprocesses sensory information is found in the retina of the eye. Here a variety of neuron types receive signals from the rods and cones (light detectors) and process information about intensity changes between neighbors (including the timing) such that they generate signals conveying information about direction of motion of objects in the visual field. Indeed the networks in the retina supply a fair amount of information that is then sent to the brain via many fewer axonal processes than one might have thought necessary. The eye produces something like primitive meaning extraction from the visual world that saves communications costs in getting signals to the brain.
Nuclei and Cortical Modules
There are two basic kinds of structures into which neural circuits are built. One is essentially a globular-like cluster called a nucleus. These nuclei are not just aggregates of undifferentiated cells or homogeneously distributed neurons. They have internal structure and may be comprised of many kinds of neurons. But they are generally more primitive in that they usually perform fairly specific functions, such as acting as distribution relays for incoming sensory data. Or they may process the data to extract meaningful spatio-temporal patterns, and if they find such patterns to signal the presence and location of objects in the environment that match those patterns. The patterns to be recognized are hard coded into the neural circuits with very little if any learning taking place. These nuclei simply respond to the presence of a specified pattern in the data stream from the sensors and activate or operate to select an evolutionarily determined response. For many of these nuclei the main output is to secrete neuromodulators that may have both a neural signaling and endocrine signaling purpose. The latter have various impacts on the physiology of the body. For example the perception of a threat may trigger a fight or flight response that includes both neuromuscular priming and prepping the body for elevated metabolism to support fighting or fleeing.
Figure 5 shows a schematic of this primitive brain architecture. This brain is designed to be a purely stimulus-response processor, having evolved from even more primitive neural networks in worms and such.
Figure 5. The earliest brains developed a consistent architecture with sensory processing feeding into several general pattern recognizing modules that were hard coded by evolution to ‘perceive’ the environment. This primitive central nervous system matched sensory data with pre-programmed reactions (affective action selection) that were then processed into motor programs for output response. Sensory inputs included body sensations that could modify the response under certain conditions. The motor programs sent signals down the spinal cord to the muscle pair groups that moved the limbs and tail (e.g. swimming fish).
Even though this kind of brain does not learn in the conventional sense, it still possesses short-term and even intermediate-term synaptic potentiation for encoding temporal traces in the circuits. It has a very primitive capacity for short-term memories so that responses that need to play out over a little longer period can do so. It would do no good to detect a threat, start to flee, and then quickly ‘forget’ the presence of danger. The beast would stop running as soon as the threat pattern was out of perception. Thus some form of memory is available in even these primitive brains, through the mechanisms described for synapses above.
Around 250 million years ago there was a major revolution in brain architecture with the development of cortical structures. These are multi-layered sheets of neurons of many different kinds that have small-scale local networks (near neighbors talking to one another) and some long-distance communications between small clusters located far away in the sheet.
Figure 6. In cortical structures one can find local clusters of tightly communicating neurons of different kinds. In this figure the red neurons send inhibitory feedback to the grey (e.g. pyramidal) cells and the blue (helper) cells. Pointed arrows are excitatory synapses and circle termini represent inhibitory synapses. Local clusters are wired to regulate the activity of the cluster. Pyramidal cells are most likely doing the bulk of long-term learning (engram encoding) whereas the other neurons are providing dynamic control, e.g. to prevent runaway excitation in closed loop feedback. Long distant communications is generally excitatory and allows the development of associations between clusters.
The main revolutionary development had to do with the cortices being organized in layers and small local clusters. The sheet arrangement enables the construction of elaborate maps, where clusters can act like positional locations of discrete percepts and concepts. Each cluster receives input from a different layer in the cortex that is receiving signals from a sensory or post-sensory processing module, such as a nucleus in the lower brain. Between the horizontal layers and the vertical clusters (e.g. cortical columns) cortical tissue appears to be well suited to encode more complex patterns than was possible in the more primitive brain. Moreover, the patterns can be learned from on-going experience rather than be hard coded into the tissues. This means that late reptiles, early mammals a birds, were able to learn important patterns in their lifetimes, meaning that they were more adaptable to changes in the environment that could never have been anticipated through evolution.
Cortices probably evolved from nuclei, some of which have layered architectures (like concentric layers in an onion). Figure 7 shows a schematic representation of an early cortical brain. Note that the old brain is still there, with some minor modifications, it is still doing its job. But now the newer cortex is added atop the old brain and receives neural as well as endocrinic (neuromodulator) signals from the older. The newer brain adds additional motor control (esp. learned behavior patterns) to the whole organism, which is relayed down through the older brain with its established interface to the musculature.
Figure 7. A major leap in the evolution of brains involved the development of cortical structures that could encode long-term memory traces of patterns and relations. The paleocortex evolved in later reptiles, birds, and early mammals. Cortices are sheets of cells in which local circuits (e.g. cortical columns) can encode object patterns (learned objects) and long-distant communications between these circuits provide ways to encode relations between objects. The cortical structures provided the first real environment learning capabilities.
The new brain is comprised of the old brain and a new cortical structure (called the paleocortex since it was evolved before the newer neocortex). The old brain still does its jobs and influences the newer cortex in doing its job. Initially this is just learning patterns and associations that help modulate the older affective pattern-response matching. To some degree the newer abilities had to regulate or modulate the older capacities in order to produce more nuanced responses to patterns that the old brain might have misclassified. But the newer brain did not yet have the more elaborate model building capabilities that would come with the evolution of the neocortex in latter mammals.
The Neocortical Brain, the Prefrontal Cortex and Executive Control
The final major revolution in brain architecture comes with the advent of a yet newer layer of cortical tissue called the neocortex. Like its predecessor it is comprised of additional layers and is internally organized to represent many more perceptual and/or conceptual details. It is also capable of forming transitory connections between clusters (now representing wholly formed concepts) to do ‘what-if’ analysis. That is, it can form hypothetical connections between dynamically represented concepts just to see what happens. This is the origin of creativity and invention. It is also the basis for the development of rational decision processing. Maps of situation concepts, decision nodes, and learned outcomes from experience can now be represented and used to guide future decisions under similar sets of conditions. This applies to learned decision paths (e.g. in expertise) and possible decision paths (models).
In order for this new capability to work, however, it needed a much more elaborate form of executive control over the forming and testing of new circuits. The motor system (or rather the premotor cortex that had evolved to do a primitive form of planning for multiple behavioral options that could be chosen as needed in highly volatile environments) gave rise to the prefrontal cortical regions in the primate brains. There we see a final convergence zone for every kind of neural signal from the most primitive parts of the brain as well as the later evolved parts. This region is responsible for thinking, but also for melding affective signals and sapient signals (judgment) with on-going rational decision processing. The latter is probably under the direct control of the premotor area (dorso-frontal area) and acts not too differently from a computer churning away through a graph search algorithm (though the analogy is not really that good - effectively the brain progressively activates concentric rings of possible clusters representing decision points. Those clusters that have strong damping either from the limbic system or from the sapience system are pruned from the decision tree. Those that have strong positive (excitation) from either system will be preferred such that the nodes they are subsequently connected to will be activated. Mutual cross inhibition of nodes further reduces the number of parallel paths as decision options are eliminated. Thinking through a problem requires both conscious directional control (such as always checking the direction against a priori desired results) and subconscious affective/judgment steering control.
Figure 8. The neocortex adds another level of complexity to the established paleocortical brain. The neocortex can encode much more detail and many more patterns than could the paleocortex. Moreover, the neocortex allowed the development of circuits that could be used to build models of the dynamics of the world as learned through experience. Much of this is encoded tacitly and can be brought to recall/usage by special executive control circuits evolved in the prefrontal cortex.
As fantastical as all of this sounds it actually represents a straightforward development of hierarchical levels of complexity based on the lower levels described above. The capacity to rapidly and tentatively form excitatory connections between cell clusters ultimately depends on the same factors that control neural wiring development (say in early embryonic development) and the dynamics of synapses able to locally modify their efficacy and maintain it for long periods of time relative to the life of the animal. These are inherent in the most primitive animal neurons. It is just minor modifications here and there in the genetic program that produces neurons and circuit forms that have been successfully selected throughout the course of animal evolution that result in the high degree of complexity we see in the human brain today.
The last step in hominid evolution involved the enlargement and capabilities of the prefrontal patch of tissue known as Brodmann area 10 (BA10). It is the patch right behind the eyebrows. It appears to have made a rapid expansion about the same time that we think humans developed language and learned to control fire. It is known to be involved in many judgment functions as well as social interactions. It is also known to have a more advanced form of communication with other parts of the brain through Von Economo (spindle) cells. Finally it is highly connected (recurrently) with every other part of the prefrontal cortex, suggesting that it is truly an ultimate convergence zone for control signals.
I posit BA10 is the seat of both consciousness (self-, other-, and of awareness) and sapience. Returning to Spock from Star Trek, it would be as if a being with only a cortical brain were possible. But this isn't the case. Our brains (and here I refer to animals in general) first evolved to monitor our bodies and our environment, to match the needs of the body and its safety with the situations found in the environment. That monitoring and control function still exists and is still carried out by the limbic and lower brain areas. But what we do know about the development of sapience is that it has high speed connections back to the limbic system that help to regulate or dampen the limbic responses. Perhaps in most humans these connections or the control competencies of BA10 are so weak that the limbic system still has undue sway over our actions. But there is some control. Rather than someone like a pure Vulcan with no limbic urges at all, perhaps the model is of a level of sapience that learns to be exceptional in down modulating emotions so that judgment can provide the better influence over intelligence and creativity. Such a capability would necessarily involve much longer range and longer time scale thinking since it would not be motivated by immediate gratification needs associated with limbic-driven decisions.
As with all but a few (r)evolutionary developments in the brain the changes have been based on simple modifications at one scale of complexity that gave rise to new competencies and even new scales of complexity. I suspect we will find the evolution of BA10 was based on some pretty minor tweaking at the genetic level dealing with the developmental control over the prefrontal cortex (you know the old story about how little difference there is between chimpanzee DNA and our own!). But because it was built on an incredible hierarchy of complexity the resulting ubercomplexification was nothing short of spectacular. I believe very strongly that with a little help, or the right evolutionary circumstances, there is more to achieve.
Very informative george altough it took ame a few days to digest!
On your section on Complexity you mention AI and I'm curious what you think of Roger Penroses ideas about 'Quantum consciousness and the possiblility that there is an essentially non-algorithmic element to human thought and human consciousness?
This non-algorithmic element to human thought is, Penrose claims, due to quantum effects in the brain, which are the source of our feelings of self-awareness, our consciousness, and our capacity for leaps of inspiration.
He also states that human mimicing consciousness in computers AI will never be achieved.
It seems plausible enough to me and the pace slow progress of work on AI suggests he may be on to something?
Posted by: GaryA | January 13, 2011 at 03:12 AM
Hi GaryA,
I've read Penrose's explanation and find it a bit too hand wavey. I do agree that consciousness is non-algorithmic but that really isn't saying a whole lot. Most processes in the real world are stochastic and heuristic (non-computable). They can, however be modeled to one degree or another as long as you inject the right kind of noise into the 'algorithms'.
The slow pace of AI I attribute to the early and continuing commitment to logic-based and algorithmic approaches. Even the so-called heuristic-based (expert systems) and experience-based (case-based reasoning) are fundamentally still just computer programs trying to explicitly compute a solution.
Neural networks and fuzzy logic have been useful in some robotic/control applications. ANNs work in pattern recognition problems quite well (e.g. the DARPA autonomous vehicle program). But the way ANNs are typically implemented they are not capable of learning on-line, real-time in a non-stationary environment. Bayesian based NNs do a little better but are still not able to learn rapidly.
Real biological neurons don't work the same way that these computationally-inspired (and implemented) systems do. If you follow the link to my paper, or check out some of my other work on my attempt to model more closely the biological neuron and synapses you will see an alternative approach. In artificial brains based on these models, the only thing being computed is the analog synaptic efficacy in multiple time domains. Lots of noise is injected to keep the stochastic flavor of real neurons. The intelligence is in the bottom up architecture of the whole system (as I tried to cover in the above).
I reached a point where I realized that further progress in this method would need a large scale simulated evolutionary environment for my 'agents', which, at the time, was beyond the capabilities of even the largest/fastest supercomputer. Today, with the parallel capabilities of the Internet, it might be feasible. Still a big project. Since I'm now trying to understand humanity's predicament I don't really have the time to devote to trying to prove this!
George
Posted by: George Mobus | January 15, 2011 at 01:18 PM
This is actually a comment sent to me by Phil Henshaw, a frequent commentator here and a blogger at synapse9.com. He reported having problems with the Typepad editor preventing his submission. If anyone else is having similar problems let me know.
------------------------------
George, As so often you do a simply marvelous job of making sense of a rather
complex subject. You also point out an important cognitive gap in your
conclusion, saying: "Real biological neurons don't work the same way that these
computationally-inspired (and implemented) systems do...". There's a rather
curious feature I noticed about that gap some time ago.
What I noticed, from air currents actually, is that they also don?t really work
the way their modeling equations do. Computational systems are all modeled
around "push" sequences (repulsions) and both thermal convection and biological
systems also have uniquely different sorts of organization that develops around
"pull" sequences (attractions). You get quite different kinds of
organizational effects from pushing and pulling on things, like from pushing and
pulling on air. One difference there is between jets of air flow versus
turbulence, sometimes characterized as the difference between pushing or pulling
on a rope.
Processes at energy gradients seem to be full of differences in organization
caused by that, and mathematical models that fail to emulate it. In systems
with many parts push relations work with global rules, but opportunistic
behaviors of energetic parts need local developmental processes to animate them
as nature does, and equations can?t be written for.
When you learn to spot the difference in organizational form, you find both in
net energy systems of all sorts. For animal populations the organization that
comes from the animation of the individuals is expressed in "foraging and
dodging". That active learning activity is what actively connects animals to
the details of their environments. You see organization arising from both the
appetites of the individuals (voluntary pull of attraction) as well as from
external forces controlling them (involuntary push of repulsion).
It's a wonderful boundary to explore, of important processes and relationships,
aided by the miss-match between the form of equations and the form of life. One
can take it negatively, as a "fly in the ointment" for representing everything
as a mechanism. One can also see it as a viewing platform for beginning
another level of exploration, of an even more richly complex natural world.
The way I use ?push models? to help me study "pull organized" things is first to
find complex things nature is doing rather simply somehow, as if following a
conventional model. That provides a reliable handle for locating and finding
the boundaries of locally animated (thought not always animate) system process.
You can then find how combinations of both kinds of order work together. It
also prompts relevant questions about the limits of such kinds of systems and
their connections. So it?s a way to use conventional models, as tools for
investigating where swarms of individual choices about their environments are
what really dominate. It seems to work.
Phil Henshaw
Posted by: George Mobus | January 27, 2011 at 09:04 AM
Hey ! am I glad to vist your blog ! from this I can get some information and facts that I didn’t know prior to. You made my day. Thank you rather much!
Posted by: Belstaff Chaquetas Venta | November 05, 2011 at 12:40 AM