NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/comp/ai/3176

Internet Message Format | 1992-08-21 | 27.4 KB

Xref: sparky comp.ai:3176 comp.theory.self-org-sys:103
Newsgroups: comp.ai,comp.theory.self-org-sys
Path: sparky!uunet!mcsun!fuug!kiae!relcom!usenet
From: [e-mail] (Vadim I. Potapov)
Subject: (1/3)REPORT: Neuroenergetic concept of intelligence
Message-ID: <[e-mail]>
Lines: 480
Sender: [e-mail]
Reply-To: [e-mail]
Organization: Financy Investment Corporation ECORAMBURS
Date: Sat, 22 Aug 92 09:42:59 +0300

NEUROENERGETIC CONCEPT OF INTELLIGENCE
L.B.Emelyanov-Yaroslavsky, V.I.Potapov
Address: Chair of Theoretical Mechanics
         Machines & Tools Institute (STANKIN)
         3A, Vadkovsky per., Moscow 101472, USSR
E-mail: [e-mail]
        [e-mail]
        [e-mail]

Abstract.
A brief description of neuroenergetic concept of intelligence
is given. The concept is based on some biological premises and
on expedient self-organization idea. The concept permits to
describe the higher function of brain in a constructive manner
and may be used both in technics and in biology.

1 Introduction
     The subject of this paper is the logic of internal
organization of a certain neuronal automaton, which permits
valuable interpretation of emotional and intellectual sides of
human brain activity. Some aspects of automaton's functioning
were simulated on computer, and obtained results confirm the
capacity for work of hypotheses supposed. Due to the lack of
place, the concept will be described in a very compressed
manner. Some hypotheses may seem controversial enough,
nevertheless such a union of hypotheses provides self-
organization and interesting integral properties of automaton,
and must be evaluated from those positions. Many properties of
automaton are borrowed from the live Nature. To emphasize this,
an element of automaton further will be called a "neuron".

2 Spike in neuron is necessary for neuron itself
     The stability of dead matter is provided by the strength of
internal bonds, while the stability of a live one is provided by
repair works. Any live unit permanently repairs itself, spending
some matter and energy. If the lack of matter and energy for
repair will take place, then a live unit will die. Such a model
of the oscillating live unit was considered in (Emelyanov-
Yaroslavsky, Potapov 1992a). In accordance with this model the
spike in a neuron is understood as an intensive repair state.
The neuron has two main properties: (1) growing old and (2) an
ability to renew itself by means of spikes. The neuron is
characterized by its age theta, which grows in the rest state
and decreases at spikes. Age theta defines feed, needed by the
neuron per unit of time (energy consumption g), and neuron's
excitability threshold P0, which is equal to external
excitation, sufficient for the transition of neuron into the
spike state. The main shape of dependencies g(theta) and
P0(theta) is shown in fig.1. One can see in fig.1, that there is
a certain critical age thetac, above which both g and
excitability of the neuron begin to increase rapidly. The neuron
has a certain internal energy reserve H, and its exhausting
involves the death of the neuron. The magnitude of H and the
rate of its variation define neuron's "self-feeling", which
influences on the excitability of a certain neuronal set by
means of "field-like" interaction.
     The neuron can stay in one of the two states: rest and
transitory state of spike. During spike the energy consumption
significantly increases, but due to decrease of theta overall
energetic effect of spikes may be positive: at a prolonged time
interval an integral energy consumption will be decreased.
External energy feed for the neuron is limited, therefore
neurons are forced to solve the consumption minimization
problem. The way of solution of this problem is the generation
of spikes, that can be expressed by the following thesis: "spike
in a neuron is necessary for the neuron itself, not for the
brain on the whole". The means of solution of this problem is
the functioning of interneuron links, creating conditions for
the mutual stimulation of spikes.

3 Optimal spike generation frequency
     Spike in the neuron is a probabilistic event: excitability
of neuron lambda0=1/P0 corresponds to the probability of a
single spike appearance per unit of time. Besides theta, lambda0
depends on other factors, and they altogether may cause
lambda0>>1, that turns neuron into the threshold element.
Conditions of spikes generation can be conveniently considered
on the plot of dependence of neuron's dynamic threshold Pd on
time tau, passed from the last spike instant. Dependence Pd(tau)
is well-known in neurophysiology and is shown in fig.2. Pd
varies in a broad limits on a very small time interval, during
which the values of other factors, influencing neuron's
excitability, practically do not vary, therefore the second and
succeeding spikes in the series of spikes may be considered as
threshold events. Spike takes place at the fulfillment of
condition W>=P0+Ps+Pd, where W is the excitation potential, that
is the result of spatio-temporal summation of influences on
neuron from excitatory and inhibitory links, and Ps is the
static threshold, which accounts for the influence of preceding
spikes. If potential W is equal to small value W1, then after
the first random spike the neuron will begin to generate spikes
with the time interval tau1, i.e. will generate at low frequency
(LF) 1/tau1 in the exaltation phase. At the great potential
value W2 the neuron will generate at high frequency (HF) 1/tau2
in the phase of deep refractoriness. As long as every spike
decreases neuron's age theta, variation of tau will vary an
energetic price of decrease of theta. It has been shown by
authors previously, that there is a certain optimal time tauopt,
which provides a minimal energetic cost. Parameters of the model
can be selected by such a way, that tauopt would be in the
region of deep refractoriness with the aim of creation of
conditions for the self-organization of the automaton: the most
natural and easy generation mode in the exaltation phase is
energetically unprofitable, while neuron can't itself reach the
frequency of generation nuopt=1/tauopt and therefore must
interact with other neurons. The tendency to the generation at
optimal frequency is provided by the optimal learning rule
(Emelyanov- Yaroslavsky, Potapov 1992a), whose central idea is
expressed by formula Drho[i][j]=nu[i](nuopt-nu[j]), where
rho[i][j] is the conductivity of an excitatory link from i-th to
j-th neuron.
     It is important for the completeness of the concept, that
besides generation at nuopt the neuron has one more
energetically profitable working mode: random spikes with
intervals, much greater than time taue, which corresponds to the
minimum of Pd.

4 Sleep and awaking
     The automaton self-organizes from an unorganized neural set
owing to the formation of a network by means of excitatory and
inhibitory links, creating conditions of mutual aid for the
solving of the problem to decrease both theta and overall energy
consumption. The first ability, appearing in the automaton, is
the ability to sleep. The sleep is a self-organization and
functioning of a big neuronal mutual aid groups (MAG's), which
form a cyclic queue for the renewal. MAG is the subset of the
neural network, which has not any internal inhibitory links. The
renewal is the HF- generation of the MAG neurons due to the
excitatory links between them, which finishes owing to the rise
of Ps as a result of generation. Due to the inhibitory links, a
strong inhibitory relations are set between MAG's, so that
working MAG does not permit to begin work for succeeding MAG,
until it stops HF- generation. As a result of the repeated
circulation of queue, the deep renewal of all neurons takes
place, i.e. theta decreases to such values, when spikes become
impossible, then the queue is destroyed and automaton switches
to the awaking state.
     In the awaking state theta rises until the moment, at
which, due to decrease of P0, the density of spontaneous spikes
becomes sufficient for the formation of a new queue of MAG's. It
is very difficult to form a queue, because at the instant of
falling asleep the energy consumption of neurons is much greater
than the feed, and the value of reserves must be sufficient for
falling asleep and for the first phase of sleep, until energy
consumption g(theta) decreases.
     Thus, the problem of automaton to exist is turned into the
problem of accumulation of reserves, sufficient for the
organization of the next sleep. Awaking state has two phases: a
phase of excessive feed at low theta values and a phase of
insufficient feed just before falling asleep. One can accumulate
valuable reserves at limited feed, i.e. to decrease overall
energy consumption even more, by prolongation of excessive feed
phase by a way of slowing down of neurons' growing old by means
of stimulation of random spikes. For such stimulation an
external excitation, i.e. the existence of receptors, is
necessary. Some results of simulation of such self-organization
in the automaton are described in (Emelyanov-Yaroslavsky,
Potapov 1992b). It is important for the further discussion, that
there exists a certain optimal value of external noisy
excitation, permitting to decrease overall energy consumption
significantly. Neuroenergetic concept of intelligence is the
presentation of intelligent and emotional properties of
automaton as epiphenomena, i.e. as a consequence of solving the
problem of search the noisy excitation sources.

5 Active memory
     It is well-known, that neural networks can realize an
associative memory function. In the neuroenergetic concept a
dynamic associative memory self-organizes in the form of
hierarchical structure of neuronal ensembles of various
modality. Neuronal ensemble is a group of neurons, interlinked
by strong excitatory links and having no inhibitory links
inside, i.e. ensemble is a MAG with good conditions for the
mutual stimulation of spikes. The first level ensembles (phoneme
ensembles, PE's) correspond to the elements of alphabets, down
to which an external situations of various modalities are being
taken apart by automaton's input devices. PE's are an immediate
source of noisy excitation for the whole neural network. Self-
organization of hierarchical associative memory is described in
(Emelyanov-Yaroslavsky, Potapov 1992b) and runs by the following
scheme:
     1. Neural network in awaking regime needs an optimal noisy
excitation flow. PE, when generating, forms one portion of noisy
flow. For the optimality of noisy flow the continuity of such
portions is necessary.
     2. The same modality PE's prevent each other to work due to
existence of inhibitory links between them. Just worked PE
creates in all others PE's the locked neurons - stoppers. Owing
to the stoppers, PE, receiving excitation from the receptor
system, can generate only at LF. To make pauses between HF-
generations of PE's small enough, it is necessary to compensate
inhibitory potentials on the stoppers by means of sending them
additional excitatory potentials.
     3. The main volume of the learned automaton memory consists
of ensembles-words (EW's), correlated with objects and
situations in the outer world. EW's are being formed when the
phoneme chains of various modalities work due to objects and
situations. Links between EW and PE's are shown on fig.3. The
main moment in such a scheme is an existence of compensating
excitatory links from EW to neurons-stoppers in PE's: by the
moment of excitatory signal arrival at the PE from receptor
devices the compensating potentials on stoppers must already
take place. This statement is in fact the statement of
foreseeing: the automaton guesses external influences, which are
to come in the near future; EW, corresponding to the object,
begins to work earlier, than the full influence from object will
come. In other words, EW is started mainly by means of
associative links from other ensembles, just have been worked:
to provide the optimal noisy flow, the automaton should work, as
a rule, in the "confirmation of expected" mode, otherwise it
will quickly fall asleep.
     4. EW is an element, having two active states. The first
state (semi-active, S) - LF-generation in the exaltation phase.
The state S is easily achievable, stable and very unprofitable:
reserves H are being exhausted rapidly. EW in the state S we'll
call the hotbed of excitation (HE). The second state (active, A)
is the HF-generation in the refractoriness phase. The state A is
a transitory one: due to the fast growth of Ps the self-locking
takes place, i.e. transition to the rest state R (A->R). The
state A is energetically profitable because of fast decrease of
theta. When i-th EW is in the state A (A[i]), the PE's of i-th
word are being started, that creates a good portion of noisy
flow, and in the automaton's memory the i-th object image is
being activated.
     5. The work of memory is necessary for memory itself: when
solving the problem of decreasing of theta for memory ensembles,
the problem of optimal noisy flow formation is being solved, and
the problem of autonomous behavior appears as a necessity to
find or construct some external situations i, providing the
activation excitation U for S[i] by scheme: S[i]+U->A[i]->R[i].
Thus, internal automaton's problem to minimize energy
consumption transforms into the problem to extinguish the
hotbeds of excitation in memory.
     6. The group of HE's {S[i]} in memory is an internal
situation in memory, which presents a concretization of
automaton's wish and compels the automaton to create an external
situation, corresponding to the internal one, i.e.
"extinguishing" it. The process of formation of HE's is the
probabilistic events of transitions R[i]->S[i], which depend on
excitabilities of EW[i]. Excitability of EW[i] can be increased
by means of increase of internal EW[i] links conductivities.
Such ensembles with raised excitability will be called the
dominants. When EW[i] achieves the state A[i], the
conductivities of internal links decrease and EW[i] ceases to be
the dominant. The dominant can be formed from any EW[i], if it
stays in the state S[i] for a long time, that corresponds to the
statement regarding the object i: "underheared", "underlooked",
"not understood". The problem of automaton's behavior in the
external environment is a reflection of its internal problem to
liquidate the dominants in memory.

6 Emotional control
     The automaton is a set of "live" units: neurons and
neuronal ensembles, which must struggle for existence, are able
to feel themselves and are in a keen competitive interrelations,
because they have the single common and limited source of feed.
For the turning of unordered neuronal set into the automaton,
one should have a sole index of state for this set and also
certain device, capable to forecast this state. The real way of
solution of this task is the introduction of a critical element
into the set, so that common motor means of the set are
controlled only by this critical element. Such a "many into
single" transformation one can imagine through the sore tooth,
when the perception is defined only by a level of toothache, and
behavior is aimed only to decrease this level. In the
neuroenergetic concept the control of associative memory is
organized on the basis of such critical element, called
emotional center (EC).
     EC is a subset of "ill" neurons, peculiarity of which is
that they cannot form links and, therefore, cannot form MAG's.
To decrease theta these neurons need regular influences from
memory, which are being formed at HF-generation of PE. The
continuous work of PE's, belonging to the same word, forms for
EC a signal, which will be called an accord. The accord compels
to work a part of neurons, having high excitability (i.e.
theta>thetac), and "overthrows" them to the low theta region,
where a neuronal batch is formed. The state of EC is defined by
the shape of its neurons theta distribution. The number of
neurons in the critical region and their positions define the EC
demand of feed, i.e. the value of feed deficit, meeting by loss
of internal reserves and directly defining the EC self-feeling.
     Feed deficit (motor tone) modulates the main processes in
the automaton: activity of memory, searching activity of
receptor systems and the magnitude of motor strength. To do a
valuable strenuous action, the automaton should carry out an
emotional mobilization, i.e. should create a big neuronal batch
in the critical region. EC self-feeling is identified with the
self-feeling of whole automaton: feelings "badly", "painfully"
are defined by value and by rate of growth of feed deficit in
EC, while feelings "well", "pleasantly" are defined by the
deficit decrease rate. The sole aim of automaton as an intact
live unit, the aim, realized at the feelings level, is the
optimization of self- feeling: maximization of "well" and
minimization of "badly". Because feelings are defined mainly by
the rate of deficit variation, then the stable states "well" and
"badly" cannot exist, and "interest to life" is supported
permanently.
     Influences on EC from working PE (accords) form on the age
distribution a compact groups of neurons, i.e. above mentioned
neuronal batches. Depending on the magnitude of accords and time
intervals between them the age distribution takes a shape of a
certain relief, which moves toward the critical region and
successively comes into it, creating the raised emotional tone
states in the automaton. Neuronal batch in the critical region,
creating a certain tonic "minus", asks the memory and external
devices to send an accord, which would overthrow the batch in
the small theta region. Thus, batches create a requests: at some
instants they demand the accords with definite magnitudes. EC is
a certain kind of dynamic rhythmic memory: a few influences are
stored in the relief, generated by themselves, and a good
agreement of newly appearing influences with the past ones,
memorized in the relief, is required further. Not any sequence
of accords can create an adequate relief in the EC, because
there are some limitations, resulting from the properties of
neurons, and defining the relationship of the sizes of batches
and intervals between them, their shape and quantity in the
period. External situations are in fact the various rhythmic
influences, and therefore they are separated onto "good" and
"bad" ones by their adequacy to the relief in EC. "Good"
situations are such that create an adequate relief easily and
quickly, and at the overthrowing of batches an emotional "plus"
appears, while the "bad" situations act in the opposite manner.
This means that for the automaton an aesthetic one is primary,
and a pragmatic one is secondary.
     At the tuning the relief in the drawing of external
influences one uses a special property of the EC neurons - they
can strongly vary their thetac value from normal position.
Variation of thetac is an internal control loop in EC, working
so that there would be neither very well nor very badly. When
the critical region contains a lot of neurons and the overall
state is bad enough, thetac moves to the right, decreasing the
maximal "badly". By means of this the automaton's hope for the
near arrival of accord and for the emotional detente is
realized. When the number of neurons in the critical region is
less than a normal value, thetac moves to the left. The velocity
of thetac movement is controlled by special memory ensembles:
the movement can be absent, and in this case the emotional
tension will grow rapidly in the presence of the batch in the
critical region, and can achieve a very great value. It was
mentioned above, that the motor tone is formed from the
emotional one, the great emotional tone is created by great
batches, penetrating deeply into the critical region, while
thetac is not moved aside. The velocity of thetac movement
becomes small at the activation of the ensemble "necessarily".
The meaning of this ensemble consists in the holding and
increasing of unpleasantness in the form of a great feed deficit
to transform it into the great motor or intellectual effort.
Thus, emotional control turns into the will control, and the
notions "necessarily", "want", "carefully" acquire a
constructive interpretation.
     The third peculiarity of EC neurons is their fast growing
old. The rate of growing old defines the velocity of relief's
movement and, hence, the internal paces of automaton and the
temporal characteristics of external situations, which can be
directly grasped by the automaton.
     From the investigation of EC point of view, of a special
interest is the "musical" problem: the obtaining of emotional
culminations on the basis of coordination of EC properties and
peculiarities of "sonic" influences to the automaton. In this
case one can hope to obtain a constructive interpretations of
mechanism, connecting the values of physiological constants with
the general peculiarities of the opuses of the "composer".
Having knowledges of this mechanism, one can try to tune up the
physiological constants in such a way, that, for example,
Tchaikovsky's music would lead to the formation of the greatest
emotional culminations, then automaton's own opuses may become
quite "tchaikovskian" ones.

7 Dynamics of EC and memory relations
     EC and memory interact so that they form together an
oscillating system. EC modulates the activity of memory by its
own self-feeling field, whose tension Q reflects the feed
deficit and influences strongly on the events of formation and
extinguishing of HE's. In the state "well" (great positive Q)
one increases the probability of events R[i]->S[i], and in the
state "badly" (great negative Q) one increases the probability
of events S[i]->A[i]->R[i]. This is provided by supposing the
dynamic threshold Pd dependent on Q: Pd(tau,Q)=Pd(tau)*eps(Q),
where eps(Q)>0 is a monotonically increasing function.
Dependence of Pd on Q is such that in the state "well" the
spontaneous excitability of ensembles is increased due to the
decrease of exaltation threshold, and at the same time
refractoriness threshold grows, while in the state "badly" the
process develops in the opposite direction. This is the basis of
mechanism of memory activity modulation by self-feeling of EC.
On the other hand, the state of the memory (i.e. the level of
"badly" for memory, defined by the number n- of HE's) influences
the value of EC neurons' thetac so that the greater is n-, the
less is thetac and, hence, the greater is energy consumption and
deficit and the worse is self-feeling.
     Such interrelations between EC and memory promote the
setting of oscillatory process: "well" in EC creates "badly" in
memory, "badly" in memory creates "badly" in EC, "badly" in EC
activates the automaton's work (perception, behavior,
recollections); all this is directed to the extinguishing of
HE's, i.e. to making of "well" in memory. "Well" in memory makes
"well" in EC: the cycle is completed and the basis for the next
cycle is formed. The automaton in the awaking state works in the
mode of such cycles-paragraphs, i.e. periods with certain
emotional and logical completeness. The scheme of a paragraph is
shown on fig.4.
     Any paragraph consists of three phases: (1) formation of
internal problem (operative aim, wish) as a group of HE's; (2)
solution of this problem - extinguishing of HE's by means of
HF-generations, and (3) evaluation of the problem - definition
of successity of its solution and correction for the future
toward the satisfiability of wishes.
     The first phase starts from the instant t0, when Q becomes
positive. At the instant t1 the first HE is being formed (in the
general case HE's may remain from preceding paragraph). At the
interval [t1,t2] the probability of formation of new HE's due to
the growth of Q is greater than possibility of extinguishing by
events S[i]->R[i], which increases with the growth of n- and
decrease of Q. At the interval [t2,t3] the number of HE's, newly
being formed, is less than being turned off ones. After the
instant t3 the events S[i]->R[i] become troubled, and the events
S[i]->A[i]->R[i], needing the activation excitation U, become
the main mean of extinguishing of HE's. The wish gets up at the
moment t3 and is concretized by the composition of HE's group
{S[i]}. The HE's group size at the instant t2 (maximal size) is
much greater than one at the instant t3, when the group
composition is relatively stable. At the interval [t2,t3] the
keen struggle of ensembles takes place for the right to remain
in a final group, i.e. to enter the wish. In this struggle such
ensembles win with a great probability, which have a good mutual
aid owing to the associative links, and which have increased
excitability (dominants).
     The second, active phase of paragraph is the interval
[t3,t4], at which the events S[i]->A[i]->R[i] take place, and
also some new S[i]'s appear due to the intensive activity of
associative search. Values of n- and Q vary insignificantly,
because, on the one hand, the new neurons come into the critical
region, but on the other hand, owing to the accords
approximately the same number of neurons decrease theta. Just
during the second phase of paragraph the needed portion of noisy
flow is formed. The meaning of this phase in the energetic sense
is the creation of conditions for HF-generation. The possibility
of HF-generation is defined by two factors: decrease of
refractoriness threshold due to the dependence of Pd on Q and
the appearing of the activation excitation U. The presence of
these factors is one of the premises for the development of
intelligence, because the latter factor connects the possibility
of automaton's energetic problem solution with the knowledge of
external environment, with the necessity of adequacy of outer
world to its internal model in the automaton. From the necessity
of approximately equal participation of both factors in the
overcoming of refractoriness threshold one should concretize the
characteristics of phoneme neurons.
     The third phase of the paragraph is its evaluation:
physiological, emotional, subjective, but reflecting in the
whole an objective evaluation of automaton's knowledges about
environment. Fig.4 shows good (solid line) and bad (dashed line)
paragraphs.  For the good paragraph at the instant t4 an
explosion of activity takes place, i.e. the simultaneous HF-
generation of a great number of ensembles, belonging the group
{S[i]} at the moment t3, owing to the associative mutual aid of
these ensembles, followed by the simultaneous extinguishing
(S[i]->A[i]->R[i]). EC receives the strong accord, removing all
the neurons from the critical (and even from pre-critical)
region. Energy consumption of EC significantly decreases, the
feed becomes excessive, Q(t) sharply grows - this is an
encouragement for the successful solution of the current
problem. For the bad paragraph the explosion of activity does
not take place either due to the weak associations or due to the
insufficient U value. As a result, the states S[i] are being
prolonged, the reserves of memory neurons are being exhausted so
that the states A[i] become impossible, there is no accord, the
emotional state is deteriorating, and at the instant t'4 the
HE's are extinguished by transitions S[i]->R[i]. Neurons in the
age distribution of EC stay in the critical region and continue
to move further - this is a penalty for the failure. At the
state "very badly" "fits" are possible, when the mechanism of
movement of thetac to the left is being turned on, the state
"badly" is being rapidly strengthened, motor tone sharply grows,
some phonemes begin to work, an accord is being formed (but
already not due to the group {S[i]} at the moment t3), EC
neurons are being removed from the critical region and a
facilitation falls.