JPH03129487A - Optimizing method for data processor - Google Patents

Abstract

PURPOSE:To escape a local minimum without fail by forcedly changing the threshold value of a neuron generating a significant output at a certain time point to be once increased to a maximum value and afterwards decreased to the value at the above mentioned time point, and adjusting weight during this period to forcedly change the threshold value. CONSTITUTION:A threshold value theta of the neuron fired by association at a certain time point t0 in a learning process is increased to a maximum value thetamax (infinitely, for example) during a following fixed period ta. The learning is continued even during this absolute inactive period and a change is applied to the weight of the neuron. Then, the neuron to be applied this change of the weight is the neuron not to be fired to the input before the t0. After the lapse of the period ta, the threshold value is gradually decreased to a value theta0 at the time point t0. In such a relative inactive period tr, the neuron not to be fired at the time point t0 is gradually ignited as well and well-balanced correction learning is gradually executed concerning a whole neural network as well. Thus, the local minimum value can be escaped without fail.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention provides a data processing system that is provided with a plurality of neurons that output data according to a comparison result between the sum of input data multiplied by a predetermined weight and a threshold value. The present invention relates to a method for optimizing the weights in a processing device.

[Conventional technology]

During the learning process of this type of data processing device, a certain degree of proper input-output correlation occurs, and when that correlation becomes stronger, the data processing device may be unable to escape from that state and optimization may become impossible. .

This is expressed as a fall to a local minimum, for example, a Boltzmann machine (David H, Ackley,
Geoffrey B., Hinton, and
Terrence J, 5ejnoWski: Al
learning a1goritb+w for Bo
lzman machines: Cognitive
5science 9.1989.147 169), we focused on the energy formula of neural networks, and in order to escape from the local local minimum of this energy formula and reach the global local minimum,
- jump the energy to a predetermined high level from time to time; However, the optimum correlation between input and output does not necessarily match the state of the global local minimum of energy. In addition, the depth of the local minimum is generally unknown, and there is no guarantee that the energy level will definitely escape from the local minimum when the energy level is increased by a predetermined value, and that the global local minimum of energy can be reached. There's not even a guarantee. Furthermore, a temporary jump in energy itself is inconsistent with the functioning of biological systems.

[Problem to be solved by the invention]

The present invention was devised to solve the above-mentioned conventional problems, and an object of the present invention is to provide a method for optimizing a data processing device that can reliably prevent a fall to a local minimum during the learning process.

[Means to solve the problem]

A method for optimizing a data processing device according to the present invention is to increase the threshold value of a neuron that has produced a significant output at a certain point in time to the maximum value while giving a constant input to the data processing device, and then increase the threshold value at that point in time. The weight is forcibly changed so as to decrease the threshold value, and the weight is adjusted during the period of this forced change of the threshold value.

Further, the method for optimizing a data processing device according to the present invention provides a smooth gradient bias to the threshold distribution of neurons in an initial state, and then, while giving a constant input to the data processing device,
The threshold of a neuron that produced a significant output at a certain point in time is forcibly changed so that it increases to a maximum value and then decreases to the value at that point, and the weight is adjusted during the period of this forced change in the threshold. It is something to do.

[Effect]

According to the method for optimizing a data processing device according to the present invention,
It is possible to escape from the local minimum value reliably by a mechanism that conforms to the function of biological systems.

〔Example〕

The present invention will be explained below with reference to illustrated embodiments.

Here, as an indicator of learning effectiveness, the following standard - number T
Let us define P.

However, here, n: number of input events i, j: number of input events Pi, j: degree of coincidence between the neural network output due to the third input event and the neural network output due to the jth human event (number of matching bits) ) The above-mentioned standard numerical value TP is an index of the discriminative ability of input events in neural networking, and the lower the TP, the higher the discriminative ability. In a neural network that has not been trained at all, TP = 100%.
, and TP monotonically decreases with subsequent learning.

(Fig. 1) However, if the learning method is inappropriate, the learning effect may be adversely affected, and as shown in Fig. 2, the TP that once decreased may rise again.

Therefore, when a drop to a local minimum occurs as described above, TP becomes saturated at a high value as shown in FIG.

This local drop to the local minimum is unrelated to the energy equation (1) above, and may be caused by, for example, bias in the weight of the features of a certain event, and a certain feature is emphasized, resulting in confusion of multiple events. This can be thought of as the result of learning to affirm .

Expressing this schematically, as shown in Figure 4, because the influence of a certain input bit 4 is strong, the firing distribution of neurons becomes biased, and this state continues to be strengthened, causing other neurons to fire. I can think of two things that make it difficult to do. In this figure, fired neurons are indicated by black circles.

In order to escape from this state, other input bits 1, ~1
．． ．． It is necessary to whiten the neuronal involvement in I.

Here, the inventors focused on the absolute refractory period and the relative refractory period in the neural network of the biological system.These refractory periods are said to contribute to the suppression and control of the propagation of sound waves in the neural network of the biological system. . (Toshi Seri: Mathematics of Neural Circuits - Information Processing Style of the Brain -; Sangyo Tosho) On the other hand, the inventors believe that these refractory periods temporarily balance the effects of each feature of the input data. Considering this as a changing factor, we changed the threshold of a neuron firing in response to a certain input as shown in Figure 5, giving the same effect as the absolute refractory period and the relative refractory period.

That is, the threshold value θ of a neuron fired by association at a certain time point to in the learning process is then increased to the maximum value θmax (for example, infinity) for a certain period of time ta. This corresponds to the absolute refractory period. Even during this absolute refractory period, learning continues and changes the weights of neurons. The neuron to which the weight change is applied at this time is a neuron that did not fire in response to that input before to. After the period ta has elapsed, the threshold value is gradually decreased to the value θ0 at time to. This period of gradual decline corresponds to the relative refractory period, and this period is designated as tr.

In the absolute refractory period, only the neurons that did not fire at time to have their weight strengthened, and learning is performed about the biased neuron distribution, but in the relative refractory period,
The neurons that fired at that point also gradually start firing, and balanced corrective learning is gradually performed for the entire neural network.

As a result, escape from the so-called local minimum is realized, and learning is performed to bring the standard device to the optimum value (global local minimum).

Note that, as in the second embodiment shown in FIG. 6, by giving the refractory period a plurality of cycles, escape from the local minimum becomes more reliable.

However, if the number of refractory period cycles is too large, the convergence of learning will be hindered, and the convergence of learning will also be inhibited when refractory period cycles are given at the end of learning.

That is, the refractory cycle needs to be given an appropriate number of times at a relatively early stage of learning.

It is also necessary to consider optimization of the absolute refractory period ta, the relative refractory period tr, and the balance between the two.

FIG. 7 shows threshold changes in the third embodiment.

In this example, no absolute refractory period is provided;
Only a relative refractory period is provided. That is, the threshold value θ takes the maximum value θmaχ at to, and then immediately decreases gradually and returns to the original value θ0 during the period ta. According to this embodiment, when there is no extreme bias in the distribution of firing neurons, it is possible to terminate learning earlier because there is no absolute refractory period.

FIG. 8 shows threshold changes in the fourth embodiment, in which a cycle of only the relative refractory period as shown in FIG. 7 is given a plurality of times.

FIG. 9 shows threshold changes in the fifth embodiment.

In this embodiment, only a relative refractory period in which the threshold value gradually increases is provided, that is, the threshold value gradually increases for a period ts from time to to reach the maximum value θmax, and immediately returns to the original value θ0. With this embodiment as well, balanced corrective learning for the entire neural network is expected.

FIG. 10 shows threshold changes in the sixth embodiment, in which a cycle of only the relative refractory period in which the threshold value shown in FIG. 9 gradually increases is given a plurality of cycles.

FIG. 11 shows threshold changes in the seventh embodiment.

In this example, a relative refractory period in which the threshold gradually increases;
A relative refractory period is provided in which the threshold value gradually decreases.

That is, the threshold value gradually increases from the time to to reach the maximum value θmax during the period ts, and then immediately decreases gradually and returns to the original value θ0 after the period Lr. This embodiment is also expected to provide the same effects as those of the embodiments described above.

FIG. 12 shows the threshold value change in the eighth embodiment, and in this embodiment, a relative refractory period in which the limit value gradually increases and a relative refractory period in which the limit value gradually decreases are repeatedly executed as shown in FIG. 1O. .

FIG. 13 shows threshold changes in the ninth embodiment. In this embodiment, no relative refractory period is provided, but only an absolute refractory period. That is, the threshold value takes the maximum value θmax for a period ta from time to, and then returns to the light value θ0. This embodiment also allows balanced corrective learning for the entire neural network in a short period of time.

FIG. 14 shows the threshold (a change) in the tenth embodiment,
In this embodiment, an absolute refractory period as shown in FIG. 13 is repeatedly executed.

FIG. 15 shows threshold changes in the eleventh embodiment. In this embodiment, there are provided a relative refractory period in which the threshold value gradually increases, an absolute refractory period, and a relative refractory period in which the threshold value gradually decreases.

That is, the threshold value gradually increases for a period ts from the time to, reaches the maximum value θmax, maintains the maximum value θmax for the period to, and then gradually decreases for a period tr, returning to the original value θ0.

This embodiment also allows balanced corrective learning for the entire neural network.

FIG. 16 shows threshold changes in the twelfth embodiment, in which a refractory period cycle as shown in FIG. 15 is repeatedly executed.

Next, in order to make the above optimization method even more effective, a description will be given of an optimization method that is implemented at the initial stage of learning.

In FIG. 17, the data processing device has a plurality of neural layers NL each having a plurality of neurons N arranged in parallel, and the neural layers are arranged so that the output of one neural layer becomes the input of the next neural layer. $, has been.

In such a configuration, the topology of the neural network can be defined, and the coordinates of each neuron can be specified as shown in FIG. 17.

Here, the X-axis is taken in the direction from the input side to the output side, and the Y-axis is taken in the width direction of each neural layer.

Then, the threshold distribution of neurons in the initial state of the data processing device is set as shown in FIG. That is, this figure shows a first example of the initial threshold value distribution. In this embodiment, the threshold values θ of neurons in the same neural layer are set to be the same, and the threshold value θ is increased as the neural layer becomes later. The gradient of increase in this threshold value is assumed to be smooth. In the normal learning process, the weight of synapses that receive significant synaptic input is increased in response to correct input;
In the threshold distribution shown in Figure 8, initially only neurons closer to the input side fire, and the range of firing gradually expands to later neurons. In this process, introducing the process of maximizing the threshold described above has the effect of increasing the number of firing neurons for a certain number of people in each neural layer, and by the time the final neural layer starts firing, many neurons are fired. neurons become involved in data processing and do not converge to a biased firing pattern. Therefore, it is possible to prevent a drop to the local minimum.

Figure 19 shows a second example of the initial threshold distribution, where:
The threshold value of the central neuron of the data processing device is increased, and the threshold value is decreased as the distance from this central neuron increases.

In such a data processing device, a firing pattern occurs around the threshold value in the initial stage of learning. However, when the threshold is maximized as described above, there are neurons with relatively low thresholds adjacent to the firing pattern, so the distribution of firing neurons gradually expands toward the top of the threshold. As a result, many neurons become involved in data processing at the end of learning, and the firing pattern does not converge to a biased one. Therefore, locals can also avoid falling into Nimam.

FIG. 20 shows a third example of the distribution of initial '#J4 thresholds, where the threshold of the neuron at the median Y coordinate is set as the lowest,
The threshold value is increased with a smooth gradient toward the maximum and minimum values of the Y coordinate. In such a data processing device, a firing pattern that passes through the vicinity of the median Y coordinate occurs in the initial stage of learning. However, when maximizing the threshold as described above, there are neurons with relatively low thresholds adjacent to the firing pattern, so the distribution of firing neurons expands toward both the maximum and minimum Y coordinates. go. As a result, many neurons are involved in data processing at the end of learning, and the firing pattern does not converge to a biased one. Therefore, a drop to the local minimum can be prevented.

[Effects of the Invention] As described above, the method for optimizing a data processing device according to the present invention increases the threshold value of a neuron that has produced a significant output at a certain point in time to the maximum value, and then decreases it to the value at the said point in time. Since the weight is forcibly changed during the period of this forced change of the threshold value, it has the excellent effect of reliably escaping the local minimum.

[Brief explanation of the drawing]

Figure 1 is a graph showing the change in standard-achievement when proper learning is carried out, Figure 2 is a graph showing the change in standard-political change when inappropriate learning is carried out, and Figure 3 is a graph showing the change in standard - political change when inappropriate learning is carried out. A graph showing the change from standard to political change when a local minimum is reached during the learning process, Figure 4 is a conceptual diagram showing an example of the firing state of a neural network that has fallen to a local minimum, and Figure 5 is a book FIG. 6 is a graph showing the threshold change according to the first embodiment of the invention method; FIG. 6 is a graph showing the threshold change according to the second embodiment; FIG. 7 is a graph showing the threshold change according to the third embodiment; FIG. Graph showing the threshold change in the example. FIG. 9 is a graph showing the threshold change in the fifth example. FIG. 1O is a graph showing the threshold change in the sixth example. FIG. 11 is a graph showing the threshold change in the seventh example. FIG. 12 is a graph showing threshold changes in the eighth embodiment. FIG. 13 is a graph showing threshold changes in the ninth embodiment. FIG. 14 is a graph showing threshold changes in the tenth embodiment. Fig. 15 is a graph showing threshold changes according to the 11th embodiment, Fig. 16 is a graph showing threshold changes according to the 12th embodiment, Fig. 17 is a conceptual diagram showing an example of the configuration of a data processing device, and Fig. 18 is an initial diagram. The graph showing the first example of the threshold distribution, Figure 19 is the graph showing the second example of the initial threshold distribution, and Figure 20 is the first example! It is a graph which shows the 3rd example of tIl threshold value distribution. Figure 70 Number of learning Figure 4 Number of learning Figure 70 Number of learning Figure 0 Number of learning Figure 1 Figure Number of learning Figure 2 Figure 70 Number of learning Figure 3 Number of learning times Figure 5 Number of learning times Figure 8 Figure 9

Claims (2)

[Claims] (1) In order to optimize the weights of a data processing device equipped with a plurality of neurons that output data according to the comparison result between the sum of products obtained by multiplying input data by a predetermined weight and a threshold value, In a method for optimizing a data processing device in which the final output of the processing device is evaluated and the weights are adjusted, the threshold value of a neuron that has produced a significant output at a certain point in time is set once while giving a constant input to the data processing device. A method for optimizing a data processing device, comprising: forcibly changing the weight so as to increase it to a maximum value and then decreasing it to the value at the point in time, and adjusting the weight during the period of the forced change of the threshold value. (2) In order to optimize the weights of a data processing device that is equipped with a plurality of neurons that output data according to the comparison result between the sum of products obtained by multiplying input data by a predetermined weight and a threshold value, In a data processing device optimization method that evaluates the final output of the processing device and adjusts the weights, a smooth gradient bias is given to the threshold distribution of neurons in the initial state, and then a constant input is applied to the data processing device. The threshold of a neuron that produced a significant output at a certain point in time is forcibly changed so that it increases to the maximum value and then decreases to the value at that point, and during the period of this forced change in threshold, the A method for optimizing a data processing device, the method comprising adjusting weights.