JPH0373057A - Rationalizing method for data processor - Google Patents

Abstract

PURPOSE:To prevent a drop into a local minimum in a learning processing by forcedly changing the weight of a neuron generating a significant output at a certain time point so that the weight can be temporarily decreased to a minimum value and afterwards increased to the value at this time point, and adjusting the weight during the period of this change. CONSTITUTION:Plural neural layers NL are provided with plural neurons N parallelly provided and the neural layer is constituted so that the output of a certain neural layer can be the input of the neural layer in the next step. Threshold values theta of the neurons in the same neural layer are mode same and the threshold value theta is increased toward the neural layer in the rear step. A gradient for this increase of the threshold value is made smooth. The weight of the neuron generating the significant output is forcedly changed so as to be temporarily decreased to the minimum value and afterwards to be increased to the value at this time point. Then, while applying the fixed input data during a period for this forced change of the weight, the weight is corrested based on the evaluated result of the output. Thus, the drop into the local minimum can be prevented.

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention provides a plurality of neurons that output data according to the result of comparing the sum of products obtained by multiplying input data by a predetermined weight and a threshold value. The present invention relates to a method for optimizing the weights in a data processing device.

[Conventional technology]

In the learning process of this type of data processing device, when a certain degree of appropriate input-output correlation occurs and that correlation is strengthened,
Thereafter, the data processing device may not be able to escape from this state, and optimization may become impossible.

This is expressed as a drop to a local minimum, such as the Portzmann machine (David) 1. Ackle
y, Geoffrey E.) Itnton, an
d Terrence J, Sejnowski:
AIearnfng algorithm for B
olztsan machines: Cognitive
e 5science 9+ 1989+ 147-16
In 9), the energy is temporarily jumped to a predetermined high level in order to escape from the energy equation of the neural network and reach the global local equation. However, the optimal correlation between input and output does not necessarily match the state of the global local energy. In addition, the depth of the local minimum is generally unknown, and there is no guarantee that when the energy level is increased by a predetermined value, it will be possible to escape from the local minimum, and the global local minimum of energy will be reached. There is no guarantee that it will even be possible. 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 it is an object of the present invention to provide a method for optimizing a data processing device that can reliably prevent a drop 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 temporarily reduce the weight of a neuron that has produced a significant output at a certain point in time to a minimum value while giving a constant input to the data processing device, and then The weight is forcibly changed so as to increase to the value , and the weight is adjusted during the period of this forced change of the weight.

[Effect]

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

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

Here, as an indicator showing the effectiveness of learning, the following standard - maturity 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 j-th input event and the neural network output due to the j-th input event (number of matching bits) ) The standard-graded TP is an index of the discrimination ability of input events in a neural network, and the lower the TP, the higher the discrimination ability. In general, in a neural network that has not undergone any learning, TP=L O0
%, and TP monotonically decreases with subsequent learning. (Figure 1) However, if the learning method is inappropriate, it may actually impede the learning effect, and as shown in Figure 2, the TP that once decreased may rise again. .

When a drop to the 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 FIG. 4, because the influence of a certain input bit I4 is strong, the firing distribution of neurons becomes biased, and this state can be considered as a state that continues to be strengthened. In this figure, neurons that have fired are indicated by black circles.

In order to escape from this state, other input bits II to I
There is a need to strengthen neuronal involvement in z, Is.

Here, the inventors focused on the absolute refractory period and relative refractory period in neural networks of biological systems.

These refractory periods are said to contribute to suppressing and controlling the propagation of excitation waves in the neural networks of biological systems. (Toshi Seri: Mathematics of Neural Networks - 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 weight 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, each weight W, ~W7 of a neuron fired by an association at a certain time to in the learning process is then set for a certain period t.
,W,i,is decreased to a minimum value,W,i,(eg, 0) during,l,. This corresponds to an absolute refractory period where the threshold value becomes infinite. 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 the neuron that did not fire in response to that input before to. After the period t3 has elapsed, each weight W to W7
is gradually increased to the value at time to.

This period of gradual increase corresponds to a relative refractory period in which the threshold value θ gradually decreases, and this period is designated as tl.

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, an escape from the so-called local minimum is realized, and learning is performed to bring the standard-political change to the optimal value (global local minimum).

Note that, as shown in FIG. 6, by applying the weight change cycle a plurality of times, escape from the local minimum becomes more reliable. However, if the number of weight change cycles is too large, the convergence of learning will be hindered, and the convergence of learning will also be hindered when weight change cycles are given at the end of the device.

In other words, the weight change cycle needs to be given an appropriate number of times at a relatively early stage of learning. Also, the period of absolute refractory period j
It is also necessary to consider optimization of the period tr of the R1 relative refractory period and the balance between the two.

Next, a data processing apparatus in which the above optimization method is effectively executed will be described.

In FIG. 7, 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 such that the output of one neural layer becomes the input of the next neural layer. It is composed of

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.

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, the threshold value θ of neurons in the same neural layer is set to be the same, and the threshold value θ is set higher 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 with significant synaptic input is increased in response to correct input, but in the threshold distribution shown in Figure 8, initially only neurons closer to the input side fire, and gradually later neurons fire. Neurons have a wider firing range.

In this process, introducing the above-mentioned weight minimization process has the effect of increasing the number of firing neurons for each 2-bilateral layer tenor constant input, and when the final stage neural layer starts firing, many neurons are fired. Neurons become involved in data processing and do not converge on a biased firing pattern. Therefore, it is possible to prevent the local name from falling into the wrong place.

FIG. 9 shows the initial threshold distribution of the second embodiment, which increases the threshold of the central neuron of the data processing device and decreases the threshold as the distance from the central neuron increases. In such a data processing device, a firing pattern occurs around the base of the threshold value in the initial stage of learning. However, when minimizing the weights described above, there are neurons with relatively low thresholds adjacent to the firing pattern, so
The distribution of firing neurons gradually widens 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, it is possible to prevent a drop to the local minimum.

Figure 10 shows a third example of the initial threshold distribution, in which the threshold of the neuron at the median Y coordinate is set to the lowest, and the threshold increases 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 the weights are minimized 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 values of the Y coordinate. 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, falling to the local minimum can be prevented.

FIG. 11 shows a fourth embodiment. This embodiment differs from the first embodiment shown in FIG. 8 in that at the same -Y coordinate, the thresholds of neurons in all neural layers are the same, while in the same neural layer, as the Y coordinate increases, the neuron thresholds are the same. The threshold value θ is increased. The gradient of increase in this threshold value is assumed to be smooth. In the threshold distribution of Fig. 11,
Initially, only neurons with large Y coordinates fire, and the range of firing gradually expands to neurons with small Y coordinates.

Here, when the weights are minimized as described above, there are neurons with relatively low thresholds adjacent to the firing pattern, so the distribution of firing neurons spreads even more smoothly.

As a result, many neurons will be involved in data processing at the end of learning, preventing convergence to a biased firing pattern and preventing a drop to a local minimum.

FIG. 12 shows a fifth embodiment, in which the thresholds of neurons in all neural layers are the same at the same Y coordinate, while at a specific Y coordinate (for example, the center) in the same neural layer. The threshold value θ in M takes the maximum value, and the threshold value θ decreases smoothly as the Y coordinate becomes further away from this value. In the threshold distribution shown in FIG. 12, neurons on the maximum and minimum Y coordinate side initially fire, and the range of firing gradually expands to neurons on the Y coordinate M side. Here, if the aforementioned weight minimization is performed, since there are neurons with relatively low thresholds adjacent to the firing pattern, the distribution of firing neurons will spread even more smoothly, and a drop to the local minimum will be prevented. Ru.

FIG. 13 shows a sixth embodiment. In this example, the threshold value θ of the neuron decreases as the X and Y coordinates become smaller, and increases smoothly as the X and Y coordinates become larger. In this threshold distribution, initially the X coordinate and Y
Neurons with smaller coordinates fire, and the firing range gradually expands to neurons with larger X and Y coordinates. Here, when the weights are minimized as described above, there are neurons with relatively low thresholds adjacent to the firing pattern, so the distribution of firing neurons spreads even more smoothly, preventing a drop to the local uniformity. be done.

FIG. 14 shows a seventh embodiment, in which the initial '#J1 threshold distribution is such that the threshold of the central neuron is the minimum value, and the threshold increases as the distance from the central neuron increases. In this example, a firing pattern is generated from the central neuron at the beginning of learning. Here, when the aforementioned weight minimization is performed, since there are neurons with relatively low thresholds adjacent to the firing pattern, the distribution of firing neurons gradually expands toward surrounding neurons. As a result, at the end of learning, many neurons become involved in data processing, preventing convergence to a biased firing pattern and preventing a drop to a local minimum.

FIG. 15 shows the eighth embodiment, where Y is the initial threshold distribution.
Regarding the coordinates, for example, the center has a high threshold value, and regarding the X coordinate, for example, the center has a low threshold value. That is, the threshold distribution exhibits a saddle shape.

This embodiment also provides the same effects as those of the above embodiments.

〔Effect of the invention〕

As described above, the method for optimizing a data processing device according to the present invention has the excellent effect of reliably escaping the local minimum.

[Brief explanation of drawings]

Figure 1 is a graph showing the change in standard-achievement when appropriate learning is performed. Figure 2 is a graph showing the change in standard-achievement when inappropriate learning is performed. Figure 3. is a graph showing the change in the standard-number position when it falls to a local minimum during the learning process, Figure 4 is a conceptual diagram showing an example of the firing state of a neural network that falls to a local minimum, and Figure 5 is A graph showing weight changes according to one embodiment of the method of the present invention, FIG. 6 a graph showing weight changes according to another embodiment, FIG. 7 a conceptual diagram showing an example of a neural network of a data processing device, and FIG. 9 is a graph showing the initial threshold distribution of the second embodiment. FIG. 10 is a graph showing the initial threshold distribution of the third embodiment. Fig. 12 is a graph showing the initial threshold distribution of the fifth embodiment, Fig. 13 is a graph showing the initial threshold distribution of the sixth embodiment, Fig. 14 is a graph showing the initial threshold distribution of the sixth embodiment. FIG. 15 is a graph showing the initial threshold distribution of the seventh embodiment. FIG. 15 is a graph showing the initial threshold distribution of the eighth embodiment.

Claims (1)

[Claims] (1) Multiple neurons are provided that output data according to the comparison result between the sum of input data multiplied by a predetermined weight and a threshold, and a smooth gradient bias is given to the threshold distribution of the neurons in the initial state. A method for optimizing a data processing device, which forcibly changes the weight of a neuron that has produced a significant output at a certain point in time so as to temporarily decrease it to a minimum value and then increase it to the value at the point in time, A method for optimizing a data processing apparatus, comprising modifying the weights based on output evaluation results while applying constant input data during the period of forced change of the weights.