Theory of cognitive regulation 2.0
A Turing machine remains the main theoretical concept
used for discussion of computations in various particular situations. Meanwhile
it is too general and for many applications it would be useful to have a
concept which is more close to human central nervous system. Theory of
automatic regulation is a mathematical expression of the principle of
self-maintenance. This principle is at the foundation of life. Stable objects
of non-living nature exist due to rigid links which keep their parts together.
In contrast, living organisms are soft, but they are capable of regeneration.
If something tries to destroy them,
they actively return themselves to the normal state. A classical regulator has
a few standard parts.
Fig. 1.
A memory cell keeps a normal value of a regulated
parameter. A sensor measures its current, real value. A comparator calculates
the difference between the 2 values. This difference passes a transfer function
and drives an executive organ which affects the input value back.
This scheme is still too general. It equally describes
operation of the nervous system and evolutionally more ancient systems which
use biochemical methods. Now let's add yet another sensor.
Fig. 2.
A classical regulator looks inside. This additional
receptor will look outside hence cognitive principle. Its destination is
initially undefined so it is essentially useless. Meanwhile this addition opens
broad horizons. Changes of the state are often the result of influence from the
environment. Using of the external effector for outside manipulations will
enhance possibilities, but without an external sensor it will be blind. It can
gauge efficiency of own operation only by the final effect on the input
parameter. A cognitive regulator can do it earlier. It also can detect various
circumstances which make operation impossible. Moreover, the external sensor
may become a proxy for the main one. It can warn about possible deterioration
in advance.
Let's apply this principle to the example of a thermostat.
The classical variant is simple. It has a work chamber, a thermometer, and 2
internal effectors - a refrigerator and a heater. Next, let's add real-world
environment. It will stand on the table near the window. Now we know the reason
of rising temperature. It is the direct sunlight. Suppose a half of the table
is in shadow. Then we will have already 2 variants of regulatory actions. We
can turn on the refrigerator or just move the device into shadow. Let's add
wheels and use a light sensor to determine the way. Live animals use their
muscles instead of wheels.
Fig. 3.
This scheme shows that while external events may
provoke some behavior, they play a modulatory role. The main trigger is inside
the organism.
Additional input-output requires more complicated
control infrastructure. One thing is clear from the beginning. A simple
transfer function is not enough anymore.
Of
course, we can invent it step-by-step by trial-and-error method, but this work
was already done by nature. The human brain is a result of long evolution and
already contains optimized solutions. Which methods has Nature found useful?
The
brain has clear layered structure. Inner parts appeared earlier in evolution.
This structure is repeated in functional architecture. This means that inner
parts remained self-sufficient. More recent additions only improved and
enhanced their operation.
When
one constructs a system of automatic regulation, the first question - which
actions to associate with deflections of particular parameters. Simple nervous
systems used the simplest solution. It is even simpler than trial-and-error
method. Deflection of a vitally important parameter triggered overall
excitation and the organism made random movements until the situation improved.
The method is very rough and the probability of success is low. Next steps of
development were taken towards more goal-oriented behavior, but in the middle
of our brain we still have a large amorphous structure called the reticular
formation. It works like the Power button of a computer - controls sleep,
awakedness, and the overall level of excitation.
The
human organism is a complicated biochemical factory, and like other
technological processes it requires thorough maintenance of certain parameters
like the body temperature, the concentration of internal liquors, etc.
Corresponding receptors and a bulk of control circuits are located in the hypothalamus.
Its output controls internal organs. Thus we have the first complete regulatory
system which is fully contained inside the body.
The
next layer has receptors directed to the outside such as hearing and vision.
Its executive organs are muscles so it operates in the outer environment too.
Its higher control center is seemingly the basal ganglia - a complex of
structures which keeps small automated elements of behavior.
Finally,
there is the neocortex - the youngest layer. It has no own receptors and
effectors. It is connected to the input-output of the previous layer. The
neocortex enhances operation exactly how universal computers improved
specialized automata - it adds programmability.
The
brain as a whole is a regulator, but like modern automatic control systems it
has got an embedded computer. That's what is interesting for us. As mentioned
in the example with a thermostat, the need in a computer appears with the
acquisition of an additional pair of external receptor-effector. The
surrounding world is much more complicated so simple regulatory circuits become
insufficient. Meanwhile, the task for this computer comes from internal
receptors. Now, such signals activate 2 pathways. An internal organ can start a
compensatory process or muscles can produce movements. In the second case,
muscular activity is usually coordinated with sensory input.
Fig.
4.
Which
parts of the brain should we include in this computer? The answer requires
understanding of some nuance about labour division between hardware and
software. In the nervous system it is different from traditional computers.
Even simplest neural nets are capable of some processing such as decision
making. If the net has an internal microstructure, this processing may be
rather complicated. It is distributed. In the live net, memory elements work
simultaneously in parallel mode, and transformation of data goes over all the
array at once. If you connect several nets together, this construct can perform
some functionality which is traditionally provided by software. The examples
include IF-THEN-ELSE, cycles, procedure calls.
As a
result, the human neocortex alone already may be considered as a
fully-functional computer. The flow of neuronal activation within it
corresponds to the flow of operators within a program. To decompile a
neuro-program, it is sufficient to trace connections between cortical areas.
The
overall picture will be as follows. Approximately a half of the neocortex
processes input from several sensory channels. Another half is involved in
generation of muscular output. Both subsystems have clear hierarchical
organization from peripheral organs to subcortical nuclei of the brain, till
the most abstract cortical fields. This structure 1:1 corresponds to event
handling in a computer. How does an event links to the corresponding handler
subroutine? There are 2 main possibilities. The first uses direct links between
sensory and motor areas. Other links go through the basal ganglia which reside
in the middle of each hemisphere. In this case, the basal ganglia work like
BIOS of a computer. They don't just relay input to output, but coordinate them
via small, system-level programs.
The
brain uses associative memory and rule-based software. Work of this
computational system is quite different from Von Neumann's processor. It has no
clock generator. Instead, circulation of activity is maintained via feedback
through the external environment. When muscles do some action, this changes the
surrounding world. These changes will be perceived and will modify conditions
of some rules. These rules will be activated and produce new actions. The
process cycles.
Such
a computational system is self-sufficient and minimally workable, but it will
be unstable. It is too dependent on external developments. The need for a
separate processor still exists. Such a processor would provide selective event
filtering and coordinated work of various links.
Copyright (c)
I. Volkov, January 27, 2014