JPH05216857A - Neural network learning state display method - Google Patents

Abstract

PURPOSE:To easily recognize a learning state by showing one quantity for showing the present learning state of each neuron, and another quantity for showing the present learning state by length of a bar, and a color or a gradation of the bar, respectively. CONSTITUTION:A display synthesizing part 4 reads out a learning state quantity from a learning state storage part 9, synthesizes plural learning state quantities designated by an operator and displays them. In this case, the display synthesizing part 4 reads out magnitude of weight from the learning state quantity storage part 9, and a converting part 4a to length of a bar converts it to length of the bar. In the same way, the display synthesizing part 4 reads out a learning pattern average of a net absolute value from the learning state quantity storage part 9, and a converting part 4b to a gradation of a color of the bar converts it to the gradation of the color of the bar. Moreover, the display synthesizing part 4 reads out a rotation angle of a weight vector from the learning state quantity storage part 9. Thereafter, these results are synthesized with regard to all neurons, and a display example is made.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to displaying a learning state in learning of a layered neural network.

[0002]

2. Description of the Related Art A layered neural network is a simplified model of a neural network in the human brain.
It is a network in which neuron neurons arranged in layers as shown in FIG. 8 are connected to neurons in other layers via synapses having a weight. Signal transmission between neurons is performed through this synapse, and various information processing can be performed by appropriately adjusting the weight of the synapse. Learning a neural network means adjusting the weights of synapses so that an expected output can be obtained.

The learning of the neural network will be described below. This is usually done using an optimization technique called backpropagation. Regarding this method, for example, PDP model, industry book (1989)
(Year) pp. 325-331.

Back propagation will be described with reference to FIGS. 1 and 9.

In step 30, the weights of the neural network 1 are initialized to start learning.

In step 34, the neural network 1 inputs the input data from the input data storage unit 7 and calculates the output of the neural network according to the following equation.

[0007]

[Equation 5]

Therefore, f is a non-linear function called a sigmoid function and has the characteristics shown in FIG.

In step 35, the learning processing unit 2 reads the teacher data from the teacher data storage unit 8 and calculates the output error of the equation 6.

[0010]

[Equation 6]

In step 36, the learning processing section 2 calculates the steepest descent gradient with respect to the error expressed by the equation (7).

[0012]

[Equation 7]

In step 32, the processes of steps 33 to 35 are repeated for all learning patterns.

In step 33, the current correction amount is obtained from the steepest descent gradient with respect to the error and the correction amount of the previous weight by using the following equation, and the corresponding weights are corrected.

[0015]

[Equation 8]

However, η is a learning coefficient, and α is a moment coefficient. The second term on the right side is called a moment term, and is an term added empirically to accelerate learning.

In step 31, steps 31 to 32 are repeated until the learning end condition is satisfied.

Problems of this learning method include (a) local minimum (b) long learning time (c) large parameter dependence of learning time.

First, (a) will be described. Backpropagation is a hill climbing method. Therefore, being caught by a local minimum is an unavoidable problem. This is a problem that depends on whether the initialization of weights is appropriate and whether η and α in Equation 4 are set properly.

Next, (b) will be described. η and α
If the value of is not set properly, learning will take a very long time.
Generally, it is very difficult to determine this optimum value.

Finally, (c) will be described. The learning time by backpropagation varies greatly depending on the values of η and α and the initialization of weights. If η and α are too small, a large number of learnings are required to reach the minimum point of the error. On the other hand, if they are too large, the vicinity of the minimum point oscillates and it does not converge easily. When the values of η and α are changed, the learning time is easily lengthened by about 10 4 as compared with the optimum case.

In the conventional learning using the pack propagation which has the above-mentioned problems, the state quantity indicating the learning process is visually displayed, and the operator monitors whether the learning is progressing properly. There is a need to.

[0023]

In the conventional display method for indicating the learning state of the neural network, the value of the weight of each synapse is shown by the gradation of color as shown in FIG.
However, when the number of neurons is large, this method makes lines that represent weights intricate and is very difficult to see. As shown in FIG. 10, even with the number of neurons at this level at most, the lines are complicated and it is difficult to grasp the state.

As another learning state display method, there is a method of sequentially showing each state quantity indicating the learning state as a function of the number of times of learning as shown in FIG. In this case, only a few data can be displayed in one graph so that the display can be easily understood. Therefore, an average is taken for each layer, or an overall average is taken to see the overall amount, and the learning state of each neuron cannot be grasped.

On the other hand, depending on the learning problem, the learning state of one neuron is poor, and the entire learning may not converge. The above display method has a problem that it is difficult for the operator to grasp such a state because the learning state of each neuron is not known.

An object of the present invention is to provide a neural network learning state display method for solving such a problem that it is difficult to grasp the learning state of each individual neuron.

[0027]

SUMMARY OF THE INVENTION The present invention provides a step of indicating, by the length of a bar, one quantity indicating the current learning state of each neuron, and another quantity indicating the current learning state of the individual neurons The learning state display method is characterized by having steps indicated by gradation.

The present invention also includes the step of indicating, by the length of a bar, one quantity indicating the current learning state of each neuron,
A learning state display method characterized by including a step of indicating a different amount of past history indicating a learning state into areas corresponding to the history and indicating each area by a different color or gradation. is there.

[0029]

In the learning state display method of the present invention, the current learning state of each neuron and the past history of the learning state quantity are visually displayed, so that the learning state of each neuron can be easily grasped. it can. Further, since the learning state of each neuron is displayed, even if the learning does not converge due to one neuron, it is possible to quickly grasp which neuron it is.

[0030]

Embodiments of the present invention will be described below with reference to FIGS.

FIG. 1 is a block diagram of the entire apparatus showing the embodiment, FIG. 2 is a flow chart of the overall processing flow of the apparatus of the embodiment of FIG. 1, and FIG. 3 is the details of the weight learning and learning state quantity calculation processing procedure. FIG. 4 is an explanatory diagram showing a display example of a learning state, and FIG. 5 is an explanatory diagram showing a display example of a history of the learning state.

Next, details of the operation of the embodiment of the present invention will be described with reference to FIG.

In step 11, the weights of the neural network 1 are initialized to start learning.

In step 13, weight learning and a learning state quantity are calculated. FIG. 3 shows details of step 13.

In step 20, the neural network 1 inputs the input data from the input data storage unit 7
Calculate the output of the neural network according to.

In step 21, the learning processing unit 2 reads the teacher data from the teacher data storage unit 8 and calculates the output error of the equation (6).

In step 22, the learning processing unit 2 calculates the steepest descent slope with respect to the error shown in the equation (7).

In step 23, the learning state quantity calculator 3 determines the net (n) of each neuron for each learning pattern.
et) is stored in the learning state quantity storage unit 9.

In step 17, the processing from step 20 to step 23 is repeated for all learning patterns.

In step 18, the present correction amount is obtained from the steepest descent gradient with respect to the error and the correction amount of the previous weight using equation 8, and each weight is corrected.

In step 19, the learning state calculation section 3 reads out the net stored in step 23 from the learning state quantity storage section 9 and also reads out the weight data from the neural network 1. Then, the learning state quantity is calculated. As the learning state quantity, for each neuron, the learning pattern average of the total input absolute value, the learning pattern average of the neuron output, the learning pattern average of the sigmoid function derivative, the weight magnitude, the error gradient magnitude, and the weight correction quantity Calculate the rotation angle of the size and weight vector. Therefore, the size of the weight,
The magnitude of the error gradient, the magnitude of the weight correction amount, and the rotation angle of the weight vector are defined by Equation 1, Equation 2, Equation 3, and Equation 4, respectively. Further, these learning state quantities are stored in the learning state quantity storage unit 9 and the learning state quantity history storage unit 10.

As a result of the above step 13, the weight learning and the learning state quantity are calculated.

In step 14, the display synthesizing unit 4 reads the learning state amount from the learning state amount storage unit 9, synthesizes a plurality of learning state amounts designated by the operator, and displays them on the display unit 6. An example of the display is shown in FIG. ◯ represents a neuron. The magnitude of the weight for each neuron is shown by the length of the bar graph above ◯, the absolute value of net is shown by the color gradation of the bar graph, and the angle of rotation of the weight vector by learning is shown by the number below ◯.

The processing for creating this display example will be described below. First, the display synthesizing unit 4 reads out the magnitude of the weight from the learning state amount storage unit 9, and the rod length conversion unit 4a converts it into the rod length. Similarly, the display synthesizing unit 4 reads the learning pattern average of the absolute value of net from the learning state amount storage unit 9, and the conversion unit 4b to the gradation of the color of the stick converts it to the gradation of the color of the stick. .. Further, the display synthesis unit 4 reads the rotation angle of the weight vector from the learning state amount storage unit 9. Then we synthesize these results for all neurons,
The display example of FIG. 4 is created.

In step 15, the display synthesizing unit 4 receives a signal from the input unit 5 as to whether or not to display the history of learning states. If a signal indicating display is received, go to step 16.

In step 16, the display synthesizing unit 4 reads the history of the learning state amount from the learning state amount history storage unit 9, synthesizes the history of the current learning state amount and the history of another learning state amount, and displays them. To display. An example of the display is shown in FIG. ◯ represents a neuron. The magnitude of the weight for each neuron at that time is indicated by the length of the bar graph above, and the past history of net absolute values up to that point is divided into areas of the bar graph and the area is colored to correspond to the history. The gradation is shown. In the figure, for example, a region showing 1.0 ≦ | net | ≦ 2.0 has a net of 1.0 ≦ | ne when the weight of the region is large.
It shows that t | ≦ 2.0.

The processing for creating this display example will be described below. First, the display synthesizing unit 4 reads the magnitude of the weight from the learning state amount storage unit 9, and the conversion unit 4a for the length of the rod converts the magnitude of the weight into the length of the rod. Similarly, the display synthesizing unit 4 reads the history data of the learning pattern average of the net absolute value from the learning state amount storage unit 9, and the bar area dividing unit 4c divides the bar area so as to correspond to the history. Further, the bar color gradation conversion unit 4b converts the net value in each area into the corresponding bar color gradation. Then, these results are combined for all neurons to create the display example of FIG.

In step 12, steps 13 to 16 are repeated until learning converges.

With such a display method, the learning state of each neuron can be grasped, and the history of the past learning state amount of each neuron can be visually observed.

In this embodiment, the learning method of the neural network is the back propagation, but the present invention is not limited to this. For example, learning using the conjugate gradient method may be used. Further, the learning state quantity is not limited to that in the present embodiment, and for example, the magnitude of the weight may be defined as in Expression 9.

[0051]

[Equation 9]

[0052]

According to the learning state display method of the present invention described above, since the learning states of individual neurons are displayed, the learning states of neurons can be individually grasped. Also, not only the size of the weight is represented by the length of the bar graph,
Since the net state is represented by the color gradation of the bar graph, it can be grasped for each neuron whether or not learning is progressing properly. This will be described with reference to FIGS. 6 and 7.

The steepest descent gradient is involved in the correction amount of the weight as shown in the equation (8). Looking at Equation 7 for the steepest descent gradient, the derivative of the sigmoid function enters in the form of a product. When the absolute value of net becomes large, the output of the neuron becomes 0 as shown in FIGS.
Alternatively, the differential value of the sigmoid function becomes small at the same time when it is saturated to 1. In that case, the steepest descent gradient becomes small, and as a result, the amount of correction of the weight also decreases, so that the learning progresses slowly. Therefore, it is possible to easily know whether or not the learning is proceeding smoothly by expressing the state of net by the color gradation of the bar graph.

On the other hand, the reason why the net becomes large is that the weight becomes large. In the calculation of net, the weight has the form of the inner product of the outputs of other neurons as shown in Equation 5, and when Net is rewritten, Equation 10 is obtained.

[0055]

[Equation 10]

From this, when the weight becomes large, n
It can be seen that the absolute value of et also tends to increase. Therefore, in order to see whether the learning is proceeding smoothly, the weight size may be displayed.

Furthermore, when learning does not converge due to one neuron, for example, when trying to recover by reducing the weight of that neuron, by displaying the history of the learning state, how small You can give guidance on what to do.

[Brief description of drawings]

FIG. 1 is a block diagram of an apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the overall flow of processing of the embodiment.

FIG. 3 is a block diagram showing details of step 13 in FIG.

FIG. 4 is an explanatory diagram showing a display example of a learning state according to the present invention.

FIG. 5 is an explanatory diagram showing a display example of a history of learning states according to the present invention.

FIG. 6 is a characteristic diagram showing a sigmoid function.

FIG. 7 is a characteristic diagram showing a differential value of a sigmoid function.

FIG. 8 is an explanatory diagram showing a structure of a layered neural network.

FIG. 9 is a flowchart showing a processing procedure of back propagation.

FIG. 10 is an explanatory diagram showing a display example 1 of a conventional learning state.

FIG. 11 is an explanatory diagram showing a display example 2 of a conventional learning state.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Neural network, 2 ... Learning processing part, 3 ... Learning state calculation part, 4 ... Multiple learning state quantity display synthesis part, 5 ...
Input unit, 6 ... Display unit, 7 ... Input data storage unit, 8 ... Teacher data storage unit, 9 ... Learning state amount storage unit, 10 ... Learning state amount history storage unit.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Junichi Higashino 1-280 Higashi Koikekubo, Kokubunji City, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd.

Claims (5)

[Claims] 1. In learning of a neural network,
Neurals characterized by the step of indicating one quantity indicating the current learning state of each neuron by the length of the bar and the step indicating another quantity indicating the current learning state by the color or gradation of the bar Network learning status display method. 2. In learning of a neural network,
The step of indicating one quantity indicating the current learning state of each neuron by the length of the bar and the past history of another quantity indicating the learning state are divided into regions corresponding to the history, and A learning state display method for a neural network, characterized in that it has a step of indicating a region with another color or gradation. 3. In learning of a neural network,
The step of indicating the magnitude of the current weight of each neuron by the length of the bar, the learning pattern average of the total input absolute value of the current neuron, the learning pattern average of the sigmoid function derivative, the output learning pattern average, and the weight vector size,
A learning state display method for a neural network, characterized in that it has a step of indicating one of the size of the steepest descent direction, the size of the correction amount of the weight, and the rotation angle of the weight vector by the color of the bar or the gradation. 4. In learning of a neural network,
The step of indicating the current weight of each neuron by the length of the bar, the learning pattern average of the total input absolute value of the neuron, the learning pattern average of the sigmoid function derivative, the output learning pattern average, and the weight vector size , The size of the steepest descent direction, the size of the weight correction amount, or the rotation angle of the weight vector, the past history is divided into areas corresponding to the history, and each area is assigned a different color or A learning state display method for a neural network, characterized by having steps indicated by gradation. 5. The weight vector size, the size in the steepest descent direction, the weight correction amount level, and the rotation angle of the weight vector in the third and fourth aspects are respectively expressed by the following equations:
A method for displaying the learning state of the neural network, which is expressed by equations 2, 3, and 4. [Equation 1] [Equation 2] [Equation 3] [Equation 4]