TOSIO KITAGAWA Research Institute of Fundamental Information Kyushu University, Fukuoka, Japan
Science,
ABSTRACT The paper deals with dynamical behaviors of the solutions of a single neuronic n-l equation: x(t+l) = l[ 2 akx(t-k)-01, where l[u] = I for u>O and = 0 for ~50, in k=O
order to have a systematic insight into a certain family of transition phenomena among 2” state configurations. For this purpose we are concerned with the specific operators of the six kinds: dpo, Zp,, ZLpcII, Yp,r, Zfll and .Z?B~(I = 0, 1, 2, , n-1). In Sections 5 and 6 we discuss reverberation phenomena for =Yz,. A regular articular representation of each state configuration is used to find out all the types and the total number of reverberation cycles under the applications of the operator LY%,,.Section 7 deals with the applications of Z’zPa,_ t, while Sections 8 and 9 deal with 3~~. Section 10 gives some ideas on the use of our family of operators for the representation of the linear threshold functions associated with the single neuronic equation.
1. INTRODUCTION The mathematical model of a neural [4-61 consists of three parts : I. neuronic II. mnemonic III. adiabatic
(or decision)
equations
(or evolution) learning
developed
by Caianiello
(NE),
equations
hypothesis
According to E. R. Caianiello explained as follows :
network
(ME), and
(or rule) (ALH). [6], the roles of these equations
are
“I. A NN (neural network) is a system composed of non-linear elements (nodes of NN) which interact through couplings (characterizing the meshes of the NN) in a manner described by NE; these give its instantaneous behavior: conventionally, a solution of the NE may be termed a “thought” of the NN. II. The strengths of the couplings may vary themselves as a consequence of the activity of the MN: such changes, when they occur, are described by ME. III. The rate of change typical of II is “secular” with respect of the times typical of I.” 0 American Elsevier Publishing Company, Inc., I973
192
TOSIO KITAGAWA
The purpose of this paper is to give some fundamental observations on the solutions of the functional equations regarding NE proposed by Caianiello [4] and discussed by him [5, 61 and his colleagues working at the Laboratorio di Cibernetica de1 C.N.R., Napoli, including de Luca [8, 91, Aiello, Burattini, and Caianiello [l], Caianiello, de Luca and Ricciardi [7], as well as many Japanese engineers and mathematicians, such as Nagumo and Sato [l 11,Amari [2, 31, Suzuki, Katsuno, and Matano [12], Ishihara [IO], and Yamaguchi [14]. The NE functional equations are given by N L(i) x,(t + 1) = 1 c c @xj(t - k) - tii
1)
j=l k=O
for i = 1, 2, . . ., N, where l[u] =
1 0 1
for u > 0, for u < 0.
(1-l) (1.2)
In Sec. 2 we shall pick up some of the main features of the functional equations given by (1 .l). Some general considerations with specific emphasis on reverberation phenomena have been given by various authors mentioned above. In contrast to these considerations, our attitude for discussing the system of functional equations is to take into consideration all the possible behavior of dynamical systems defined by (1.1). Due to this attitude, a systematic description of the dynamical systems is indispensable, and we shall discuss not only each individual graph but also a family of graphs in which transitions among graphs are one of our main concerns. In accordance with such a research approach we shall propose to discuss a certain family of transition phenomena: (i) convergences (ii) convergences (iii) disintegration figurations.
to stable configurations, to reverberation of reverberation
cycles, and cycles into
other
types
of con-
We are also interested in a detailed discussion of various types of reverberation cycles. We shall, therefore, introduce several notions that are useful in discussing the dynamical behavior of the solutions of the functional equations. It is to be noted here that, although our ultimate goal is to give a set of universal observations on the dynamical behavior of the solutions of functional equations including (1. l), we shall start with a specific type of the equation : n-l x(t + 1) = 1 c a,x(t - k) - E) (1.3) [ k=O which is an NE for a single neuron.
1,
ON NEURONIC
EQUATIONS
193
It is indeed our approach to start with the simplest equation and to discuss all the possible features of the dynamical behavior of all possible solutions of the functional equations and then to proceed to more complicated equations. By proceeding in this way, we are expecting to prepare a transition map of various configurations. Such a transition map will give us meaningful observations on evolution equations and learning processes when we enter into the consideration of ME and ALH in the sense of Caianiello [4, 61. In Sec. 2, characteristic features of the neuronic functional equations are summarized from mathematical points of view. Since, in this paper, we confine ourselves to a single neuronic equation (1.3), there is a certain set of conventions that simplify the notation given in Sec. 3. Section 3 also explains the principal attitude of our research in three assertions (a), (b), and (c) with an illustration of the special case of (1.3) when n - 1 = 2. Section 4 is devoted to the introduction of concurrence functions and allied operators defined over the whole space of 2” state configurations. We are interested in various features of the transition of state configurations under sequential applications of each of these operators. In this paper we are concerned the specific operators of the six kinds: pw, Tz, Zm,, Ye,, Yp,, and 2~~ where I = 0, 1, 2, . . ., n - 1. In Sec. 5, reverberation cycles for Ztp,, are discussed. It is immediately clear that any state configuration belongs to one and only one reverberation cycle under applications of Ya,. Proposition 5.2 shows that any state configuration has a regular articular representation by which it can be directly observed that it belongs to a reverberation cycle of specific length. A regular articular representation of each state configuration is also useful to find all the types and the total number of reverberation cycles for n-dimensional state configurations under application of the operator Ta,. It is shown in Proposition 6.10 that the total number N(Z) of reverberation cycles having the exact length I can be obtained through a certain set of difference operators. Section 7 deals with reverberation cycles of n-dimensional state configurations under applications of $Pd,,_,. In this case there exists a set of state configurations that are not a member of any reverberation cycle but are transitory state configurations in the sense that repeated applications of Y,,_, will carry each of them into some state configuration, which is a member of some reverberation cycle called the #“‘*-reverberation cycle. Here again an articular representation is equally important as in the case of Za,, in Sec. 6, however, with more delicate complications. In Sets. 8 and 9 we turn to a family of transition operators called 2& (I = 0, 1, 2, . . .) n - 1). Preparatory considerations on Yp, show that a reverberation cycle under application of YE0 can be disintegrated
194
TOSIO KITAGAWA
into a set of state configurations consisting of transitory state configurations and those belonging to a reverberation cycle(s) under application of dpp,. It is noted that S,,, and YZ, can be considered to be Y;“aOand Zp,, respectively. We proceed to introduce Yppl (1 = 2, 3,. . ., n - l), and some introductory observations on the mutual relationship among these operators are given with reference to simple examples. In order to have a summarized feature associated with transition of state configurations induced by application of Y,-,, we shall introduce a new coordinate system suited to Yp, for each fixed 1, where 1 s I 5 n - 1. Dynamical behavior induced by application of each YF~_ j (j = 1, 2, 3, 4, 5) is discussed in some detail. After these preparations, we turn, in Sec. 9, to a discussion of reverberation cycles under application of P’p,, where the uses of the new coordinate system introduced in Sec. 8 play an important role. Throughout these eight Sections we have been substantially concerned with the specific family of operators PO, Z’,, P’,,, YZo, Yb,, and 28, (I = 1, 2, . . .) n - 1). In fact, although our explanations have been limited to Yii, and ZD,, situations are simpler for the cases of Ya and PG, and analogous observations are valid between P’@, (= YD,,) and ZMO (= Yb,), and also between Ts, and 55’~~respectively. At this stage we wish to make clear the conditions imposed on the set of coefficients 0 or even the {a,_,_,> (k = 0, 1,2, . f ., n - 1) and the parameter function 0(t) by which the translatable functional operator defined in (1.3) becomes equivalent to one member of the operator family consisting of J!?~, YG, YO,, and Yp, (I = 0, 1, 2, . . ., n - 1). This topic will be discussed in a future paper [l]. It will be shown that the uses of the operator family have a certain significance provided that specific considerations are given in composing these operators. In this paper, Sec. 10, an illustration of the systematic uses of the operator family is given with regard to the special case of (1.3) when n - 1 = 1. It is also noted that there is a need for introducing the notion of conditional operator representations of l[z@)]. A few remarks on the change of operator representations in connection with the functional equation when 0 in (1.3) is a function of t. 2. CHARACTERISTIC FEATURES FUNCTIONAL EQUATIONS
OF THE
NEURONIC
values Let xi(t) (i = 1, 2, . . ., IV) be N functions of all the nonnegative oft which takes either 1 or 0 for each t. We assume that a set of functional equations (1.1) hold where {a!$)} (i,j = 1, 2, . . ., N; k = 1, 2, . . ., Y(i)) of t. and {Oi> (i = 1, 2, . . ., N) are assigned sets of constants independent
ON NEURONIC
195
EQUATIONS
The characteristic features of a system of functional can be summarized in the following five aspects.
equations
(1.1)
(1) Threshold functions induced by the function (1.2) involve nonlinearity of the functional equations and also reduce discrete activities of N neurons. (2) Networks equations (1.1).
among
N neurons
are defined by a system of functional
(3) Translations with time are our concern with respect to a system of functional equations (I) which defines a dynamical system of Ndimensional vector X(t) = [xi(t) x2(t), . . ., x~(t)]. (4) The functional
operator
I Y(t) defined by
JJ,(t + 1) - Al { Y(f)> I{ Y(t)} =
y,(t + 1) - My(t))
)
. . . . . . . . .
with f&{r(t)} = 1
[
i
;
k=O
j=l
a$‘yj(t
- k) - ei
translatable, i.e. commutative with all translations where T,Z(t) E Z(t + c(), in the sense that we have T,(I{ Y(t)>) = IV,
(2.1)
>
( .JG + 1) - MWI
1)
(2.2)
T,, - 00 < c(< 00,
Y(t))
(2.3)
for all t and c(. (5) Transition map among 2N(L+ l) state configurations the system of functional equations (1.1).
are defined
by
In combining these five characteristic features of the system of functional equations (1.1) we observe that we are now concerned with a dynamical system of the Markovian type under a formulation of finite mathematics. 3. SINGLE
NEURONIC
EQUATION
In what follows we confine ourselves to a single neuronic equation where the number of neurons in Eq. (1 .l) reduces to one. However, for the sake of brevity of notation, we rewrite our equation, without loss of generality, in the following form: II-1
1)
1) = 1 c a,x(t - k) - 0 (3.1) [ k=O which is a special case of (1) essentially with N = 1 and L = n - 1. Now let us introduce an n-dimensional vector x(t
+
6, = (60 ht-1, 6,-z, . . .YL_(n--2), &-(“-I)),
(3.2)
196
TOSIO
KITAGAWA
and hence the inner product n-1 (a, 6) = ,zO GL!f. Our concurrence
function
(3.3)
~(6,) is a real-valued
function
z(6,) = (a, 6,) - 0. The functional We may and vectors defined by
equation
(3.1) is now reduced
6 f+l = shall introduce
Mm
by (3.4)
to
= f(h).
a transformation
defined
9
(3.5) of n-dimensional
4 4-l 8
(3.6)
Ez.
L-3) t - (n -
1;
In view of the translatability property given in Sec. 2, we may and we shall confine our discussion to the transition
(3.7)
Now 6, is called to be a state at the time point t of our dynamic system, state while 6, = [6,, 6f--l, 81--2, . . ., 6t_cn-1J is called an n-dimensional configuration at the time point t of our dynamical system. The functional equation (3.1) induces a digraph (X, I) where X is the set of all 2” state configurations, that is, X = (6) with 1x1 = 2”, where the directed connection I, that is, the set of arcs (edges), is defined by the map I- : 6 -+ Lq6). (3.8) The principal three assertions:
attitude
of our approach
(a) We set up a digraph but
(b) We are interested also with transition
(X, I) induced
is summarized
by the following
by (3.8).
not only in each individual digraph given by (a), phenomena among these digraphs which are
ON NEURONIC
EQUATIONS
197
induced
by the changes of Y(6) due (a,, a1 2 . . .) a,_,, e>. (c) The dynamical behavior of equation as t + + cc is our concern, to rest (critical) state configurations, problems associated with them.
to the changes
of the vector (a, 0) =
the solution (6,) of the functional including the studies of convergence reverberation cycles, and the stability
The following examples are given in order to illustrate general problems with which we are concerned in this paper. In order to simplify an intuitive understanding of the fundamental features introduced here, we shall appeal to the notations specially suited to each individual example. EXAMPLE
3.1
Let us consider x(t + I) = I[a,x(t)
+ a,x(t
-
1) + a,x(t
- 2) - O]
(3.9)
which is a special case of (3.1) with z(a, 6) = C&)6, + a,& + a,& -Or; 0
Notation:
- 8.
ao12 I no +a1 I-n2 aij 5
ai -hi,
(Odi.jS2,
i+j)
FIG. I. Transition map associated with (3.9).
(3.10)
198
TOSIO KITAGAWA
A transition map is shown in Fig. 1, where to each route there is associated the condition under which the route is realized. The transition map includes all the combinations of possible routes which can be realized under any feasible conditions upon the set of coefficients (a,, a,, a2) and 0. For the moment we write z(6,, 6,,6,)
= a,&
+ a,&
+ a,&
There are eight state configurations (Table corresponds a function value of z(6,, a,, 6,).
0
0
0
-0
001 0 1
0
ai -
e e
1
1
al
a2
0
a2
-
+
-
fJ
(3.11)
- 0.
1) to each of which there
1 0
0
a#J -
1 1
1 0 1
da + a2 -
1
0 1 1
0
uo+ala0
+
al
i-
0 0 a2 -
0
We consider a set of the following five cases (I-V) each of which is defined by a combination of two sets of conditions upon z(6,, 6,, S,), respectively. There is a set of common conditions which each of all these cases satisfies, and also another set of conditions which one and only one of these cases satisfies, respectively. Now the common conditions are given: (1) z(6,, 6,, 1) 5 0