JPH0424756A - Learning machine - Google Patents

Abstract

PURPOSE:To reduce the learning time of a learning machine by dynamically selecting weight change direction from plural directions and changing the weight, and dynamically setting a learning parameter and an acceleration parameter where the error becomes the least. CONSTITUTION:An input signal is weighted and added at a hidden layer 22 and an output layer 22, an output signal is obtained by executing a non-linear processing, and the change direction of the weigh is decided at a weight change direction decision part to minimize the error between a desirable output signal outputted by an instruction signal generation part 13 and the output signal of the output layer. After that, the weight change amount to the plural learning parameter is obtained by a straight line location part 67, and while checking whether the learning is not in an inefficient state by an unsuitable learning state detection part, each weight is changed by adding a fixed amount times the weight change amount of the last time to the weight change amount to the learning parameter with the minimum error obtained by the straight line location part 67 at a weight change part 65. Thus, the error is made sufficiently small by repeating to the change direction of the weight. Thus, the error becomes sufficiently small in a short learning time, and the learning is ended.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a learning machine for a data processing device.

Conventional technology As a conventional learning machine, 4L, for example, "Learning" by E. Rummelhart et al. '(Lea
rning representativeons b
y back-propagating erro
rs)', Nature Vol.
323 No. 9 (1986). 1st O
The figure shows a general configuration diagram of a conventional learning machine.
201.202 are input terminals, 217 is the same learning area 21&
219. 220 is a multi-input-output circuit 221 is an output layer 222 is a hidden layer. As shown in Figure 10, the learning machine processes signals input from the input terminals using a configuration in which eight power circuits are connected in a hierarchical manner.
Output from the output terminal. In this way, a board of multi-input-output circuits connected in a hierarchical manner.The layer consisting of multi-input-output circuits that output output signals is called the output layer, and the other layers consisting of multi-input-output circuits are hidden layers. The hidden layer may be composed of multiple input-output circuits forming one layer, or may be composed of multiple input-output circuits forming multiple layers. Figure 11 shows the conventional learning method. The configuration diagram of the machine is shown. In FIG. 11, 201.202 are input terminals, 203. 204
．． 205, 20a, 207. 208 is a variable weight multiplier 20 and 21o.

211 is an output terminal of an adder 212 having saturation input/output characteristics, 213 is a teacher signal generation section 214, an error calculation section 215, a steepest descent direction determining section 216 is a weight change a 217 is a learning same section, 218.219.220 is a multi-input terminal -Output same area 22
1 is the output @ 222 is the hidden layer Multi-input-output circuit 218, 219 and 2 as shown in Figure 11
20 consists of a variable weight multiplier and an adder with saturation input/output characteristics. Immediately, the output signal of the j-th multi-input-output circuit is y[j co - fnc(Σ(W[lT j co zero y
[i ko))-----(1) is expressed as a here E y[
x] is the output signal of the first multi-input-output circuit in the previous layer, and w[i, j] is the output signal of the first multi-input-output circuit in the previous layer. This is the weight applied when manually inputting data to a multi-input-output circuit. fncO is a function with saturation characteristics and is expressed as a sigmoid function or the like.

FIG. 12 shows a graph of the characteristic function of the adders 20Q, 210, and 211 having saturation input/output characteristics expressed by fncO. A learning machine has a structure in which multiple input-output circuits like this are connected in a hierarchical manner, and during learning, variable weight multiplication is performed to output a desired output signal (hereinafter referred to as a teacher signal) for the input signal. Vessel 203
．． 204.205.206.207 and 208 are changed. To find the amount of weight change
First, find the teacher signal, the output signal of the output layer, and the convergence error = E (w) (2). Here, GQ ya[J] is the output signal ts[j] of the j-th multi-input-output circuit of the output layer for the 9th input signal.
is the teacher signal for ye[j] Σ is the general rule for all the teacher signals Σ is the general rule for all the teacher signals in the output layer General summary for multi-input-output circuits W is the weight w[i.

j] (hereinafter W is called a weight vector). As shown in equation (2), the error E is expressed as the sum of squares of the difference between the teacher signal and the output signal of the output layer. In learning, the weight is changed to minimize the base, or error, between the teacher signal and the actual output signal, but the amount of weight change is E △ W −−ε 0 +
α book Δ W ° ... (3) & W where ε is a positive constant called learning parameter α is a positive constant called momentum, and W is the error expressed by equation (2). A vector whose component is the differential of E by the weight w[i, j], and is called the direction of steepest descent. ΔW′ζ This is a vector representation of the amount of weight change in the previous learning.

13 shows a configuration diagram of the learning circuit 217 of this conventional learning machine. In FIG. 13, 223 is an input terminal for output layer output, 224 is an input terminal for hidden layer output, 225 is an input terminal for input signals, 226 is an output terminal for output layer weights,
227 is an output terminal for hidden layer weights. In the learning circuit 217 of the conventional learning machine, the teacher signal generator 213 generates the teacher signal (desired output signal) to[j] for the input signal.
occurs. On the error calculation unit 214i Teacher signal L[j
] and the output signal y-[j] of the output layer, and the error E expressed by equation (2) is calculated. The error calculation unit 214 outputs a difference signal te[j] −ya[j] between the teacher signal and the output signal necessary for changing the weight to the steepest descent direction determining unit 215. The steepest descent direction determination unit 215 determines the steepest descent direction to the error E in the weight space expressing the weights as vectors based on the difference signal, the output layer output signal, the hidden layer specific force signal, the human input signal, and the weight of the output layer. The steepest descent direction is determined by the right side of equation (4), which is a vector expression of the differential by the weight of the error E.
4 & Multiply the direction of steepest descent by the learning parameter and output to the weight changing unit 216. The weight change unit 216 calculates the weight change amount using equation (3). Each variable type multiplier 203,
204.205. 20a, 207 and 208
As described above, by repeatedly calculating the amount of weight change using the steepest descent method, the error becomes smaller, and when the error becomes small enough, the output signal becomes sufficiently close to the desired value. As a matter of fact, the study is completed.

FIG. 14 shows a detailed block diagram of the learning circuit 217 of a conventional learning machine. In FIG. 14, 228 is a learning completion determination K, 229 and 230 are differential coefficient calculation units, 231 is an output layer differential coefficient storage unit 232, a δ storage unit 233 is a δ*W buffer, and 234 is a hidden layer differential coefficient storage unit 235 is an input Signal memory ff1L 23a, 237.23 & 239
and 240 are multipliers 241.24 and 243 and 244
is an adder 245 and 246 are learning parameter multipliers 2
Reference numeral 47 denotes an output layer weight storage section 248 that is a hidden layer weight storage section. The operation of the learning circuit 217 will be explained in detail below using FIG. 14. The amount of weight change output by the steepest descent direction determination unit 215 is expressed as where w[i, j] is the third multi-input - the i-th multi-input of the previous layer input to the output circuit - The weight applied to the output signal of the output circuit is 6w, [i, j] is the amount of change in the direction of steepest descent of w[i, j], and ε is a learning parameter. (5
) expression is transformed...(6) where nets[j ko − Σ(w[i, j ko book yo
[x], and ys[i] is the output signal of the i-th multi-input-output circuit in the previous layer for the ninth human input signal. Equation (6) is further modified as △ws[i, j] = ε book Σ (δ
w*[j komoto yo[i]) ... (7) However, below is a modification of the equation (yo) When the third multi-input-output circuit is in the output layer, and when the third multi-input -Different when the output circuit is in the hidden layer.

・It can be expressed as when the third multi-input-output circuit is in the output layer. However, fnc(netp[j]) is the characteristic function (= yp[j]) of each multi-input-output circuit in the output layer. Using equation (2), this equation can be further transformed and finally expressed as (9).In FIG. 14, the adder 241
calculates (tp[j]-yo[j) in equation (9),
is calculated. Since this value is δwp[j], it is stored in the δ storage unit 232. The multiplier 237 multiplies the product of δW, cj: ] and the hidden layer output y, [i] by the bundle number. The multiplier 245 multiplies the learning parameter ε to obtain Δw+[
i, j] is calculated by adding α times the weight change amount in the previous learning to △wa[i, j] in equation (7) to obtain the weight change amount △w[i , j] Adder 2
Calculate the sum with the weight before change in 43. Change the output layer weight stored in the output layer weight storage unit 247. 6w when the third multi-input-output circuit is in the hidden layer.
s[j BE δnets [J] However, neto[kl−Σ(w[j, kl y−[jl)
Here, w[j, kl is the weight applied when the output y*[jl of the third multi-input-output circuit is input to the multi-input-output circuit in the next layer. Using equation (8), at t1,
This equation is further transformed and becomes...(10). In FIG. 14, the multiplier 238 outputs the output δwe[kl of the δ storage unit 232 and the output layer weight w[j, kl
Find the product of δ*W buffer 233 and adder 242 to find Σ(δw=[k]kyo w[j, kl). The multiplier 23 via the differential coefficient calculation unit 23 and the differential coefficient storage unit 234
9 and said Σ(δw=[k] books w [j, kl
) can be used to find δwo[jl in equation (10).
The multiplier 240 multiplies the input signal yp[i] and the multiplier 24
In step 6, △wJi,j], which is expressed by the equation (a), is obtained by multiplying the learning parameter ε by the equation (A). △w a [i
, jl is multiplied by α times the weight change amount in the previous learning, and the weight change amount ΔW [i, jl is a bundle number expressed by equation (3). Number Changes the hidden layer weight stored in the hidden layer weight storage unit 248.

As described above, in order to reduce the error between the output signal of the output layer and the teacher signal, the difference signal between the output signal and the teacher signal is calculated. Find the amount of weight change sequentially. This learning algorithm is called the error backpropagation method, but the problem that the invention aims to solve is that the learning parameter ε
Since fixed values are used that are determined empirically or found through trial and error, they are not necessarily optimal values and have the problem of increasing the time required for learning, and changing the weights. The amount (Y) is a value obtained by adding a certain percentage of the previous weight change amount to a certain percentage of the differential by the weight of the error E, which has the problem that the amount of weight change is not necessarily optimal. There is a problem that learning falls into an inefficient state in which the error does not decrease even if the process is advanced.The present invention takes this into consideration and minimizes the error between the teacher signal (desired output signal) and the actual output signal. Another object of the present invention is to provide a learning machine that requires less time for learning by dynamically setting the direction that minimizes the error while proceeding with learning. The objective is to provide a learning machine that requires less time for learning by dynamically setting the optimal value of the learning parameter and proceeding with learning while taking into account the amount of previous weight change.

Means for Solving the Problems The present invention provides a hidden layer consisting of a plurality of multi-input-output circuits that perform non-linear processing on a weighted sum of human input signals using a characteristic function having saturation characteristics and output the resultant signal, and an output signal of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing on a weighted sum using a characteristic function having saturation characteristics and outputs the result, a teacher signal generation section that generates a desired output signal of the output layer, and a teacher signal generation section of the teacher signal generation section. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the output and the output of the output layer; and outputs weight change amounts for the plurality of learning parameters regarding the weight change direction. The learning machine is equipped with a straight line search unit that changes the weight by adding a constant times the previous weight change amount to the weight change amount found by the straight line search unit.

According to the above-described configuration, the present invention weights and adds human input signals in the hidden layer and the output layer, performs nonlinear processing, and obtains an output signal.A desired output signal outputted by the teacher signal generation section and an output signal of the output layer are obtained. In order to minimize the error between The weight change unit changes each weight by adding a constant times the previous weight change amount to the weight change amount for the learning parameter that minimizes the error found in the straight line search unit. The same goes for the rest. Repeat the operation of changing the weights using the learning parameters that minimize the error in the weight change direction to make the error sufficiently small. As described above, once the weight change direction is determined, the optimal learning parameters are dynamically set in that direction, and the weights are changed by adding a constant times the previous weight change amount to the weight change amount for the optimal learning parameter. By making Ct small, the error becomes sufficiently small and learning can be completed in a short learning time. Decide the direction of change by selecting from multiple directions. For example, the multiple weight change directions may be selected from the direction of steepest descent and the direction of conjugate gradient. It is locally the most efficient direction in the sense that the amount of error decreases the most.On the other hand, the conjugate gradient direction changes the weights in a direction different from the direction in which the weights were changed in the previous learning. In the sense of
This is the most efficient direction from a broader perspective. therefore,
In the present invention, the time required for learning is shortened by selecting and dynamically determining the weight change direction from among these multiple weight change directions and using the optimal weight change direction.

Embodiment FIG. 1 shows a configuration diagram of a learning machine in a first embodiment of the present invention.a In FIG. 1, 1 and 2 are input terminals, &4, 5 . a, 7 and 8 are variable weight multipliers 10
and 11 is an output terminal of an adder I2 having saturation input/output characteristics, 13 is a teacher signal generator 14 is an error calculation surface 15 is a steepest descent direction determination expansion 21 is a pressure [22 is a hidden limit 65 is a weight change unit 66 is a learning The parameter initialization unit 67 is a straight line search unit 93
94.95 and 96 are multiple input-output circuits.

The learning machine of this embodiment has a configuration in which multiple input-output circuits consisting of variable weight multipliers and adders with saturation input/output characteristics are connected in a hierarchical manner. -The output circuit weights and adds the signals, and performs nonlinear processing using the characteristic function of each adder to obtain the output signal.

In the learning of the learning machine of this embodiment (table), the weights multiplied by the variable weight multipliers 3, 4, 5, 6. If the vector whose components are these weights is called a weight vector, the amount of change in the weight vector can be expressed as a vector.The direction of this weight change vector is called the weight change direction. ) Since the error E is a function of the weight vector, it can be expressed as a curved surface in the weight space where the weights of variable weight multipliers 3, 4, 5, 6, 7 and 8 are expressed as vectors. This is called the error surface J%. Figure 2 shows a schematic diagram of the iso-abandon line of the error surface showing the operation of this embodiment.
In the figure, w[i+, j+] and w[i2. ja
] is the weight of any two variable weight multipliers among variable weight multipliers 3.4.5.6.7 and 8=7hw− is a vector representation of the weight value initialized with a random number g. P is the minimum error point in the straight line search in the direction of gs in the first learning, and gl is the steepest descent force KP2 in Pl is the minimum error point in the straight line search in the direction of g+ in the second learning. P2 is a point indicating the value of the weight after the second learning. In the learning of the learning machine of this example (Table 1), first the weights of all variable weight multipliers are initialized with random numbers. The purpose of learning is to start from the starting point in Figure 2 and repeat the search for the minimum error point by changing the weights in the direction of decreasing the error on the error surface until the global minimum error point is reached. In the first learning, the steepest descent direction g- found by equation (4) is used as the weight change direction and Lg.
Find the point where the error is minimum in the direction of g
The learning parameter initialization unit 66 outputs an appropriate positive value as the initial value of the learning parameter in the straight line search. Show the diagram. In FIG. 3, ε1 is the learning parameter initialization unit 66
is the initial value of the learning parameter set by ε k −ε k−1 pieces 2 (k>1)
...(13), and Em (k≧0) is a false base for the learning parameter ε, that is, E= −E (zero g・ for w・+ε) (k≧0.
ε・20)...(14), and εV is the learning parameter tEν at the vertex of the parabola that approximates the error surface.
is the error for the learning parameter εν. Figure 4 is a PAD showing the operation of the learning machine of this example in a straight line search.
Diagram (Problem Analysis Diagram)
m). As shown in FIG. 4, in the straight line search, first find the error with respect to the initial value ε1 of the learning parameter. The straight line search unit 67 multiplies the initial value ε1 of the learning parameter by the direction of steepest descent, and outputs the weight change amount for the initial value ε1 of the learning parameter to the bundle number weight changing unit 65. The weight change unit 65 calculates the sum of the weight change amount and the weight and outputs the weight for the initial value ε1 of the learning parameter.

An output signal is determined in the output layer 21 using these weights, and an error is determined in the error calculation section 14 by comparing the teacher signal and the output signal. In this way, the initial value ε of the learning parameter
The initial value E1 of the error relative to 1 is determined. P in Figure 4
As shown in the AD diagram, the straight line search unit 67 compares the initial error value E1 with the error E at the starting point of the error surface, and calculates E.

If Ct becomes smaller than force <, E-, it can be expected that the error for a learning parameter with a large value will be smaller. Repeat until it turns. In the first learning shown in Figure 3, El < E-, so by repeating the operation of doubling the value of the learning parameter ε and calculating the error, the number becomes E・>El> Ee < Es.
The straight line search unit 67 also calculates the weight change amount for ε3.
Note that the error El force (initial value E of the error) with respect to the initial value of the learning parameter.

When ζ increases, we can expect that the error for learning parameters with smaller values will be smaller.
The straight line search unit 67 repeats the operation of multiplying the learning parameter value by 1/2 until the error becomes smaller than the initial error value E-. Next, in order to approximate the vicinity of the minimum error point with a parabola, the straight line search unit 67 finds errors at points where the weight values change at equal intervals in the weight space. Water saving εa, s − (ε2 + ε3)/2 (15), ε2 − ε11 ε26 − ε2 ε3 − ε2.5
The error for ε2.6 is E2.6 − E(w + εa, s zero g
@) ...(1'6) is also found. In Figure 3, fi E2.6 < Ee < Es〈El, so
The error surface in the vicinity of the minimum error point is approximated by a parabola passing through the three points where the error is minimum among these points, and the error at the vertex is determined. The weight change amount is calculated by multiplying the learning parameter at the vertex of the parabolic approximation of the error surface by the bundle number, steepest descent direction ga, and ε using the node plate.The weight change unit 65 sets the weight value for the learning parameter εV, and the output layer 21, the output signal is calculated, and the error calculating section 14 calculates the error EV.
Ea, E2. The weight change amount ε2.5 for the learning parameter ε2.5 which compares s and Es and gives the smallest error E2.5 among them. The weight change unit 65 calculates and outputs the weight change amount εe, 511g.
The sum of m, a constant times the previous weight change amount, and the weight W- is calculated and stored as a new weight.

In the first weight change, the previous weight change amount is zero, so even if the weight is changed to Wl Mi We + ε260g・, the weight change amount in the first learning is △w +
The above is the first learning. In the second learning, the steepest descent direction g1 at the point P1 represented by Wl in the weight space is determined by the steepest descent direction determination unit 15 by the bundle number. The learning parameter initialization unit 66 (Y) sets the initial values of the learning parameters to ε1 and L for the first learning, and the direction of the learning parameters used for the previous weight change for the second and subsequent learning. value or ε
The larger value of 1 is determined. In learning after the second time (friend) If the optimal value of the learning parameter from the previous time is thicker than ε1, the optimal value of the previous learning parameter becomes the initial value of the learning parameter; When it is small, ε1 becomes the initial value of the learning parameter.4 As a result, the optimal value of the previous learning parameter is used as the initial value of the learning parameter for the current learning, and the optimal learning parameter for the current learning can be set efficiently. At the same time, when the optimal value of the previous learning parameter is small, use it as the initial value of the learning parameter for the current minimum point search to reach the local minimum point of the error surface and reach the global minimum point of the error surface. However, in the straight line search unit 67, the learning parameter can be doubled or 1 in the direction of gl.
Multiply by 72, approximate the error curve to a parabola near the weight where the error is small, find the weight where the error is minimum in the direction of gl, and find the minimum error point P for the second learning. Also, if the value of the learning parameter used at this time that minimizes the error is εopt, the weights will be W2"Wl 10 ε o@t by the second learning.
gl + α book △ Changed to Wl. However, α is a positive constant called momentum, and a
The amount of weight change in the second learning is △ We −ε 6−t 0 g 1 +
Even if it is stored as α zero △ Wl, the weight changes are repeated in the same way until the error becomes sufficiently small. In this way, the learning machine of this embodiment advances learning by repeatedly searching for a straight line in the direction of steepest descent.
The optimal value of the learning parameter is dynamically set to reduce the error and approach the global minimum error point.

As described above, according to this embodiment, the learning parameters that minimize the error are dynamically set by straight line search. By adding a constant multiple of the previous weight change amount to the weight change amount that minimizes the error, the calculation is performed based on not only the local shape information of the error surface but also the more global shape information of the error surface.
Compared to the case where a constant times the previous weight change amount is not added to the weight change amount that would cause the largest error,
Even if the error is minimized by changing the weights a small number of times, the learning parameter initialization unit 66 (Y) in this embodiment sets the initial value of the learning parameter to ε1 for the first learning and L2 for the second and subsequent learnings. The strength is determined to be the larger value of the optimum value of the parameter or ε1.The initial value of the learning parameter may always be set to a fixed value ε1.
,X.

In this embodiment, the hidden layer 22 is composed of a multi-input-output circuit that processes human signals and outputs them to the output layer.However, it may also be composed of multi-input-output circuits connected in a hierarchical manner. In this embodiment, a force outputs one output signal for two input signals (the number of these input/output signals can be any number). This is a diagram showing the configuration of the machine. In FIG.
Reference numeral 92 indicates an unsuitable learning state detection circuit 97, which is a learning circuit.The difference between this embodiment and the first embodiment is that in the first embodiment, the direction of weight change is the direction of steepest descent, whereas in this embodiment In the example, the point is that the direction is selected from two directions, the steepest descent direction and the conjugate gradient direction. In this embodiment, the hidden layer sum of products storage section 77 and the unsuitable learning state detection section 92 are newly provided. The figure shows a configuration diagram of the learning circuit 97 of this embodiment.

In FIG. 6, 23 is an input terminal for output layer output, 24 is an input terminal for hidden layer output, 25 is an input terminal for input signals, 2
6 is an output terminal for output layer weights, 27 is an output terminal for hidden layer weights, 73 and 74 are adders, 75 is a weight memory ii
84 is an output terminal of the hidden layer output, 85, 86 and 87 are multipliers 88 are the product-sum storage unit 89 of the input signal and the change direction, and the product-sum storage unit 89 is the product-sum storage a of the input signal and the weight. 90 is an adder. The old one is nonlinear conversion. Department. Learning circuit 9 of the learning machine of this embodiment
Table 7 and 6 The teacher signal generator 13 gives a desired signal as a teacher signal as the output of the output layer for the human input signal, and the error calculation unit 14 outputs a difference signal between the teacher signal and the actual output signal.Determine the conjugate gradient direction. Section 16 (Y) Based on this difference signal, the direction of weight change is determined to be the direction of steepest descent or the direction of conjugate gradient. Line search section 67 determines the amount of weight change for a plurality of learning parameters as a bundle number. Hidden layer sum-of-products storage section 77
is the output signal of the hidden layer for each learning parameter, and the bundle number error calculation unit 14 calculates the error for each learning parameter as the bundle number.
The weight change unit 65 changes the weight using the learning parameter for the smallest error among these errors. If a larger value is detected, it is determined that the learning has reached a state where the error cannot be efficiently reduced even if learning is continued further, and the weights are reinitialized with random numbers and learning is resumed.

In this embodiment, as shown in FIG. 6, the hidden layer sum-of-products storage unit 77 is used to obtain the output signal of the hidden layer.

The output signal yt[j of the hidden layer for the input signal yp[i]
l is yp[jl = fnc(Σ(w[i, j
It is expressed as l zero y-[11))-(18). Just L
w[i, jl is the weight multiplied when the human input signal yp[i] is input to the third multi-input-output circuit in the hidden layer, 1), fncm is the characteristic function of the adder in the hidden layer, and It is a nonlinear function with characteristics.

w of the weight change direction d determined by the conjugate gradient direction determination unit 16
The component corresponding to the amount of change in [i, jl is d[i, jl
Expressed as, the hidden layer output signal ye[jl for the learning parameter ε is ye[jl = fnc(Σ
ε zero d [i, jl) zero yJx co))...(1
9) Therefore, is it necessary to perform multiplication twice as many times as the number of human input signals each time yp[jl is calculated?a In this example, (YoA e
[J co = Σ (w[i, jl customer yp[i co)
---- (20) is stored in the product-sum storage unit 8 of the input signal and the weight.
By storing Bp[jl −Σ (d[i, jl yp[i ko) (21) in the product-sum storage unit 88 of the input signal and the change direction, Hidden layer output signal ys E j for learning parameter ε
] to yJj] fnc(Σ ((w[i, j + ε td [i, jl) Book 3'f[i])) fnc
(Σ (w [i, j] ys [i ko) ten ε zero Σ (d [i, j] wealth yp [i])) - fnc (Ae
[j]+ε Ning B p[jl)...-(22). Buried board Product-sum storage unit 8 for input signal and change direction
The multiplier 87 multiplies the output Be[j(of
Find the book Bp[jl). In the nonlinear transformation unit 91, the characteristic function fnc() of the adder of the hidden layer is applied to the output of the adder 90.
A nonlinear transformation equal to L is applied to y expressed by equation (22)
Find a[J]. As a result, the number of calculations can be significantly reduced in the straight line search in the weight change direction d,
The learning time can be shortened.

In this example, the conjugate gradient direction is used as the direction in which the main weights are changed. , β is a constant given by g゛12, d' is the conjugate gradient direction in the previous learning, lK17'l is the norm of the vector in the direction of steepest descent g' in the previous learning, and weight change in the first learning. Direction (main: Determine the direction of steepest descent.

FIG. 7 shows a detailed block diagram of the conjugate gradient direction determination unit 16. In FIG. 7, 28 is the steepest descent direction calculation unit 29 is the conjugate gradient direction calculation unit 30 is the weight change direction determination unit 31 is the learning number counting unit 32 is the input terminal for the weight of the output layer, and 33 is the weight change direction of the output layer. , 34 is an output terminal for the weight change direction of the hidden layer, 35 is an input terminal for the difference signal, 36 and 37 are for calculating differential coefficients, 38 is for output layer differential coefficient storage a, 39
40 is the hidden layer differential coefficient memory 40 is the δ memory 41 is the current output layer's steepest descent direction storage section 42 is the previous output layer's steepest descent direction memory a 43 is the β calculation unit 44 is the previous output layer weight change direction Decision unit 45, 46, 47, 48, 49, 50 and 51 are multipliers, 52 is an adder, 53 is a δ*W buffer,
54 is the input signal storage R, 55 is the steepest descent direction storage unit 56 of the current hidden layer, and the steepest descent direction storage unit 5 of the previous hidden layer.
8 is the previous hidden layer weight change direction memory @ 59 and 6
0 is weight change direction switching K. 63 and 64 are adders.

The operation of the conjugate gradient direction determination unit 16 will be explained in detail with reference to FIG.
The multiplier 46 outputs the direction of steepest descent regarding the weight of the output layer, and the multiplier 50 outputs the direction of steepest descent regarding the weight of the hidden layer. These signals representing the direction of steepest descent (i.e., the weight change direction switching units 59 and 60 of the weight change direction determining unit 30, the steepest descent direction storage unit 41 of the current output layer of the conjugate gradient direction calculation unit 29, and the current hidden layer) The steepest descent direction in the current learning stored in the steepest descent direction storage section 41 of the current output layer and the steepest descent direction storage section 55 of the current hidden layer is output to the steepest descent direction storage section 55 of the current output layer. The β calculation unit 43 calculates β based on the steepest descent direction in the previous learning stored in the output layer steepest descent direction storage unit 42 and the previous hidden layer steepest descent direction storage unit 56. Multiplier 48 and 51
is the product of the weight change direction d' of the previous learning and β.The adders 63 and 64 calculate the sum with the direction of steepest descent, and the weight change direction switching unit 59 of the weight change direction determining unit 30 determines the conjugate gradient direction. and output to 60. As described above, the weight change direction switching units 59 and 60 are manually operated in the steepest descent direction and conjugate gradient direction. In this learning, a selection signal is output so that the direction of steepest descent is the direction of weight change.In this example, there are six weights multiplied by variable weight multipliers 3, 4, 5, 6.7, and 8. , A selection signal is output so that the steepest descent direction is set as the weight change direction in learning once every six times, and the conjugate gradient direction is set as the weight change direction in other times of learning.According to this selection signal, the weight change direction is set as the direction of weight change. Switching section 5
9 and 60 switch the weight change direction between the steepest descent direction and the conjugate gradient direction. As described above, the conjugate gradient direction determination unit 16 determines the weight change direction.
In step 2, when it is detected that learning has fallen into an inefficient state,
By re-adjusting the weights and restarting learning, learning is prevented from falling into an inefficient state and the time required for learning is shortened. By storing the sum and the product sum of the input signal and the weight change direction of the hidden layer, the number of calculations in the line search for the weight change direction d can be significantly reduced, and the learning time can be shortened. . According to this embodiment, the straight line search unit 67 calculates the amount of weight change for a plurality of learning parameters in a bundle, and the weight change unit 65 calculates the weight by adding a constant times the previous weight change amount to the weight change amount that minimizes the error. By changing the weights, it is possible to dynamically set the optimal learning parameters for the direction of weight change and change the weights.By changing the weights based on more global information of the error surface, The time required for learning can be shortened. Also, the time required for learning can be shortened by determining the conjugate gradient direction or steepest descent direction as the weight change direction in the conjugate gradient direction determination unit 16. The conjugate gradient direction determining unit 16 sets the weight change direction to the determined conjugate gradient direction or the steepest descent direction (the changing direction may always be the steepest descent direction). The direction of change is set to the direction of steepest descent when learning is performed once every time equal to the number of weights, and is set to the conjugate gradient direction during learning other times.When the change direction is learned once every half of the number of weights, it is set to the direction of steepest descent. For other times of learning, the conjugate gradient direction may be used (-Also, in a straight line search for the conjugate gradient direction, if the error does not decrease even if the weight change amount is smaller than a certain value, the weight change direction may be set to the steepest descent direction. Also, in this embodiment, the inappropriate learning detection unit 92 detects that the error is in an inappropriate state where the error cannot be efficiently reduced even if the learning is further advanced by using a certain absolute value of the weight. The force is detected by re-initializing the weights with random numbers and restarting learning (you can adjust the weights by uniformly compressing all weights, or compress only the weight with the largest absolute value) Also, in this embodiment, the hidden layer 22 consists of a multi-input-output circuit that processes human input signals and outputs them to the output layer. In this embodiment, one output signal is output for one or two input signals.The number of these input/output signals may be any number.

FIG. 8 shows a configuration diagram of a learning machine according to a third embodiment of the present invention. In FIG. 8, reference numeral 68 denotes a momentum straight line search unit 98, which is a learning circuit. The feature of this embodiment is that after performing a straight line search in the direction of the steepest descent, a straight line search is performed in the direction of the previous weight change ((The direction of the weight change is dynamically set at the same time as the magnitude of the weight change.

FIG. 9 is a schematic diagram of the iso-abandon line of the error surface showing the operation of this embodiment. In Fig. 9, △W1 is the weight change in the first learning, IP+ is the error maximum iJ of the straight line search regarding the steepest descent direction in the first learning, asgl is the steepest descent force K at Pl, and P2° is g in the second learning.
The minimum error A PR of the straight line search in the direction of l is P2'
The direction of weight change in the first learning from (direction of △W1)
In the learning circuit 98 of this embodiment, the weight change amount △w1 is determined by a straight line search regarding the steepest descent direction at the starting point. First, perform a straight line search regarding the direction of steepest descent g1 at Pl (\ Find the point P2° with the minimum error. Next, Pa
A straight line search is performed from ' in the direction of the previous weight change amount ΔW1 (
\ Find the point P2 with the minimum error.The value of the learning parameter that results in the minimum error with respect to the direction of steepest descent used at this time is ε.

If αopt is the momentum value that minimizes the error in the direction from Bt, p2' to △W1, the weight will be W2''Wl + ε oot by the second learning.
+ α o 11t zero Δ Even if it is changed to Wl, the amount of weight change in the second learning is △W2” εoptgg+ + αapt Book
It is stored as ΔW1. To explain the above operation with reference to FIG.
7 outputs the amount of weight change for a plurality of learning parameters regarding the steepest descent direction, the momentum straight line search unit 68 sets the momentum to zero, and the hidden layer 22 and output layer 2
1 to find the error for these weight change amounts. Find a point P2' on the error surface for the weight change amount that is the minimum among these errors. Straight line search unit 67 outputs the weight change amount corresponding to P2' L Momentum straight line search unit 6
8 adds α times 6w1 (α is momentum) to the weight change amount and outputs the result to the weight change unit 65, and the hidden layer 22 and output layer 21 calculate an error with respect to the momentum.

The momentum straight line search unit 68 sets a plurality of momentums and finds the momentum that minimizes the error among those momentums. The weight change unit 65 changes the weight by the weight change amount for the momentum with the minimum error.The weight change unit 65 repeats the weight change in the same manner until the error becomes sufficiently small. In this way, the learning machine of this example dynamically sets the optimal values of the learning parameters and momentum by repeating the straight line search for the steepest descent direction and the straight line search for the previous weight change direction. Reduce the error.

As described above, according to this embodiment, the learning parameters and momentum that minimize the error are dynamically set by the straight line search. Conventional learning machine Compared to , the error is minimized by changing the weights a small number of times.

In this embodiment, the hidden layer 22 is composed of a Takuker output circuit that processes human input signals and outputs them to the output layer.
It may also be configured with multiple input-output circuits connected in a hierarchical manner. In addition, in this embodiment (1 for the two input signals in the table)
The number of these input/output signals may be any number.

As described in detail, according to the present invention, it is possible to dynamically select the weight change direction from a plurality of directions to change the weights, and dynamically set the learning parameters and acceleration parameters that minimize the error. , the weight can be changed by adding a fixed ratio or an optimal ratio of the weight change amount in the previous learning to the learning parameter weight change amount with the minimum error, and the learning time of the learning machine can be shortened.

[Brief explanation of the drawing]

Fig. 1 shows the configuration of the learning machine according to the first embodiment of the present invention. Fig. 2 shows the operation of the same embodiment and is a schematic diagram of the iso-abandoned line of the error surface. Fig. 3 shows the starting point of the error surface of the same embodiment. Cross section between point and Pl @Figure 4 shows PADI! which shows the operation in straight line search of this embodiment! FIG. 5 shows the configuration of the learning machine according to the second embodiment of the present invention. FIG. 6 shows the configuration of the learning circuit 97 of the second embodiment. FIG. FIG. 8 shows the configuration of a learning machine according to the third embodiment of the present invention. FIG. 9 shows a schematic diagram of the iso-line of the error surface showing the operation of the third embodiment. FIG. 10 shows the general configuration of a conventional learning machine. @ Fig. 11 shows the configuration of the conventional learning machine. Fig. 12 shows the adder 20λ 210 and 2 of the conventional example.
FIG. 13 shows the configuration of the learning circuit 217 of the conventional learning machine. FIG. 14 is a detailed block diagram of the learning circuit 217 of the conventional learning machine. 15... Steepest descent direction determination wind 16... Conjugate gradient direction determining unit 65... Weight changing unit 67... Straight line search unit 77... Hidden layer sum of products memorization 92... ... Name of agent for detecting inappropriate learning state Patent attorney Shigetaka Awano 1 person Figure wl ” Wb ” ” 1 Figure 1 W[+2 jz+ J) 1 2 y[j l 130

Claims (14)

[Claims] (1) A hidden layer consisting of multiple input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
A learning machine comprising: a weight change unit that changes the weight by adding a constant times the previous weight change amount to the weight change amount found by the straight line search unit. (2) A hidden layer consisting of multiple input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
a weight change unit that changes the weight by adding a constant times the previous weight change amount to the weight change amount found by the straight line search unit; 1. A learning machine comprising: a hidden layer sum-of-products storage unit that stores sums of products with weight change directions. (3) A hidden layer consisting of multiple input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
A weight change unit that changes the weight by adding a constant multiple of the previous weight change amount to the weight change amount found by the straight line search unit, and an inappropriate state in which the error is not efficiently reduced even if learning is further advanced. What is claimed is: 1. A learning machine comprising: an inappropriate learning state detecting unit that detects that the learning machine is falling into. (4) A hidden layer consisting of a plurality of multi-input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals and outputs the result, and a weighted sum of the output signals of the hidden layer that has a saturation characteristic. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
a weight change unit that changes the weight by adding a constant times the previous weight change amount to the weight change amount found by the straight line search unit; A hidden layer sum-of-products storage unit that stores the sum of products with the weight change direction, and an inappropriate learning state detection unit that detects that the error is in an inappropriate state where the error does not decrease efficiently even if learning continues. A learning machine characterized by: (5) A hidden layer consisting of a plurality of multi-input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals and outputs the result, and a weighted sum of the output signals of the hidden layer that has a saturation characteristic. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
A learning machine comprising: a weight changing unit that changes the weight by dynamically setting a ratio of a previous weight change amount to the weight change amount found by the straight line search unit. (6) A hidden layer consisting of a plurality of multi-input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
a weight change unit that changes the weight by dynamically setting a ratio of the previous weight change amount to the weight change amount obtained by the straight line search unit; and a sum of products of the input signal and the weight of the hidden layer, and the input signal. 1. A learning machine comprising: a hidden layer sum-of-products storage unit that stores sum-of-products of the weights of the hidden layer and the changing direction of the weights of the hidden layer. (7) A hidden layer consisting of a plurality of multi-input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
a weight change unit that dynamically sets the ratio of the previous weight change amount to the weight change amount found by the straight line search unit, and changes the weight; A learning machine comprising: an inappropriate learning state detection unit that detects that the learning machine is in an appropriate state. (8) A hidden layer consisting of a plurality of multi-input-output circuits that performs nonlinear processing using a characteristic function that has a saturation characteristic on the weighted sum of input signals, and has a saturation characteristic on the weighted sum of the output signals of the hidden layer. an output layer consisting of a multi-input-output circuit that performs nonlinear processing using a characteristic function and outputs the output; a teacher signal generation section that generates a desired output signal of the output layer; an output of the teacher signal generation section; and an output of the output layer. a weight change direction determining unit that selects and determines a weight change direction from a plurality of directions to minimize the error between the weight change direction and a straight line search unit that outputs weight change amounts for a plurality of learning parameters regarding the weight change direction;
a weight change unit that changes the weight by dynamically setting a ratio of the previous weight change amount to the weight change amount obtained by the straight line search unit; and a sum of products of the input signal and the weight of the hidden layer, and the input signal. A hidden layer sum-of-products storage unit that stores the sum of products with the changing direction of the weight of the hidden layer, and inappropriate learning that detects when the error is in an inappropriate state where the error cannot be efficiently reduced even if learning continues. A learning machine characterized by comprising a state detection section. (9) A weight change direction determining unit that determines a weight change direction to be the steepest descent direction or the conjugate gradient direction for minimizing the error between the output of the teacher signal generation unit and the output of the output layer. The learning machine according to any one of claims 1 to 8. (10) The present invention further comprises a weight change direction determination unit that determines a weight change direction to be the direction of steepest descent for minimizing the error between the output of the teacher signal generation unit and the output of the output layer. 9. The learning machine according to any one of 8 to 8. (11) An inappropriate learning state detection unit is provided which detects, by determining that the absolute value of the weight is larger than a certain value, that the error is in an inappropriate state in which the error is not efficiently reduced even if the learning is further progressed. Claims 3, 4, 7,
8. The learning machine according to any one of 8, 9, or 10. (12) If the error is not efficiently reduced even if learning continues, it is detected that the absolute value of the weight is larger than a certain value, and the weight is re-initialized with a random number. The learning machine according to any one of claims 3, 4, 7, 8, 9, or 10, further comprising an inappropriate learning state detecting unit that restarts learning. (13) If the error is not efficiently reduced even if learning continues, it is detected by the absolute value of the weight being larger than a certain value, and all weights are uniformly compressed. The learning machine according to any one of claims 3, 4, 7, 8, 9, or 10, further comprising an inappropriate learning state detection unit that restarts learning after the learning has been completed. (14) An inappropriate state in which the error cannot be efficiently reduced even if learning is continued is detected by the absolute value of the weight being larger than a certain value, and the absolute value of the weight is the largest. Claim 3, further comprising an inappropriate learning state detection unit that uniformly compresses only the weights and restarts learning.
, 4, 7, 8, 9 or 10.