JP2006178666A - Information processor, information processing method and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an information processor capable of corresponding to a plurality of pairs of a plurality of input/output relations in operation simulating a neural network, and to provide also an information processing method and a recording medium. <P>SOLUTION: The information processor provided with an external input terminal 2 for receiving an external input signal, an external output terminal 4 for outputting an external output signal to the external and a plurality of operating elements 8a-8f simulating nerve cells is also provided with an independent switching part 5 having a plurality of switches 5a-5c for switching external input signals inputted from the external input signal 2 to the plurality of operating elements 8a-8f and output signals from the operating elements 8a-8f and capable of independently driving the plurality of switches 5a-5c. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an information processing apparatus, an information processing method, and a recording medium for performing an operation simulating a neural network.

  An information processing apparatus that performs an operation that simulates a neural network is configured by constructing a large number of unit operation elements that simulate a nerve cell by learning and combining the many unit operation elements. Such an information processing apparatus conventionally includes an input layer that receives an input signal, an output layer that outputs an output signal, and an arithmetic unit including a large number of unit arithmetic elements. Only the relationship between the two output signals could be learned. Therefore, in this information processing apparatus, when one input signal is input, one output signal can be obtained.

In addition, the information processing device further includes an input reproduction layer that reproduces the input signal. The output of the input reproduction layer is fed back to the input layer after fixing the output of the output layer, thereby reproducing the input signal from the output signal. (See Patent Document 1). In this information processing apparatus, it is possible to learn about a relationship in which a plurality of output signals correspond to one input signal and a relationship in which one output signal corresponds to a plurality of input signals. Therefore, in this information processing apparatus, when one input signal is input, a plurality of output signals can be obtained, and when a plurality of input signals are input, one output signal can be obtained.
Japanese Patent No. 3180334

  Since the information processing apparatus feeds back the output of the input reproduction layer to the input layer, learning can be performed only when the state of the system is uniquely determined by the input signal. Therefore, it is possible to learn about a relationship in which a plurality of input signals correspond to one output signal and a relationship in which a plurality of output signals correspond to one input signal, but a plurality of output signals with respect to a plurality of input signals. I can't learn about the relationship that corresponds to. That is, learning is possible when one of a plurality of input signals or a plurality of output signals is designated in advance, but learning is performed when a plurality of input signals or a plurality of output signals are not designated in advance. I can't. Accordingly, a plurality of output signals can be obtained from one input signal, and one output signal can be obtained from a plurality of input signals, but a plurality of output signals cannot be obtained from a plurality of input signals.

  An information processing apparatus that can perform an operation that simulates a neural network is applied in various fields. However, the application range is limited unless a plurality of output signals are obtained for a plurality of input signals. As a result, the application range is narrowed. For example, when applied to a control device that outputs a control signal for driving a plurality of actuators based on detection signals of a plurality of sensors, it must cope with various situations such as failure of some sensors and actuators. Therefore, it is necessary to learn a relationship in which a plurality of output signals (control signals) correspond to a plurality of input signals (detection signals). In addition, when applied to a device that communicates with people, there are many-to-many relationships such as the names of (cedar, pine, beech) and the concept of (wood, timber). Since there are many of them, it is necessary to learn a plurality of relationships related to each other.

  Therefore, an object of the present invention is to provide an information processing apparatus, an information processing method, and a recording medium that can handle a plurality of input / output relationships in a calculation that simulates a neural network.

  An information processing apparatus according to the present invention is an information processing apparatus including an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells. A plurality of switches for switching an input signal to the arithmetic element between an external input signal from an external input terminal and an output signal from the arithmetic element, and the switch includes an independent switching unit that operates independently. And

  In this information processing apparatus, the external input signal from the external input terminal and the output signal (feedback signal) from the arithmetic element are independently switched by each switch in the independent switching unit, and the external input signal and the output signal of the arithmetic element are switched. One of the signals is input as an input signal of a plurality of arithmetic elements. Since the input / output can be arbitrarily switched in this way, the information processing apparatus can learn a relationship in which a plurality of external output signals correspond to a plurality of external input signals. In addition, in the information processing apparatus, when an arithmetic element that learns a plurality of input / output relationships is provided, a plurality of external output signals can be obtained for a plurality of external input signals. Data can be retrieved.

  The information processing apparatus according to the present invention may include a calculation state determination unit that determines a processing state in the information processing apparatus based on an input signal to the calculation element and an output signal from the calculation element.

  In this information processing apparatus, an input signal to the arithmetic element (the output signal from the arithmetic element is a feedback signal that is fed back through the independent switching unit and is the previous output signal) by the arithmetic state determination unit and the arithmetic element The processing state is determined based on the output signal. That is, the time change of the signal output from the arithmetic element is monitored, and it is determined whether the output signal has changed or has converged. By determining the processing state in this way, it is possible to determine whether or not the signal output from the arithmetic element has converged to a certain value. As a result, the information processing apparatus can output only a signal whose value is determined as an output as an external output signal.

  The information processing apparatus according to the present invention may be configured to output an external output signal even when there is a contradiction between the external input signal and the external output signal.

  This information processing apparatus outputs an external output signal even when the external output signal is inconsistent with a certain external input signal (when there is no mutual relationship between the signals). For the purpose of idea support and the like, it is preferable to output the interrelation between signals including inconsistencies that are not related to each other between the external input signal and the external output signal. In other words, various ideas arise from contradictions, so signals that contain contradictions are also important outputs. From such contradiction, it is possible to construct a new neural network that has an idea that cannot be considered by common sense.

  The information processing apparatus according to the present invention may be configured to clearly indicate a portion having a contradiction.

  In this information processing apparatus, the inconsistent part is clearly indicated, and the inconsistent part is notified to the user. As an explicit method, for example, when there is a monitor, display using the color of the contradictory portion, change in contrast, change in texture, change in font, etc., use change in audio if there is audio output. .

  The information processing apparatus according to the present invention may include an external input control unit that performs control so that external input signals are sequentially input to the arithmetic elements.

  In this information processing apparatus, the input order of a plurality of external input signals to the arithmetic elements is controlled by the external input control unit, and the sequentially input external input signals are sequentially input as input signals of the arithmetic elements by the independent switching unit. When there are a plurality of external input signals, different outputs may be obtained by changing the input order. Therefore, it is possible to narrow down information in an arbitrary order by controlling the input order. As a result, an interface having a conversation function that enables human interpretation can be realized.

  An information processing apparatus according to the present invention is an information processing apparatus including an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells. An error calculator that calculates an error based on an input signal to the element and an output signal from the arithmetic element, and the error calculated by the error calculator along the propagation direction of the signal from the input signal to the output signal of the arithmetic element A forward learning unit that propagates and adjusts the arithmetic element so as to reduce the error.

  In this information processing apparatus, an error calculation unit calculates an error between an input signal and an output signal with respect to an arithmetic element. In the information processing apparatus, an error is propagated along the direction in which the signal flows from the input signal to the output signal by the forward learning unit, and the arithmetic element is adjusted so that the error is reduced. As described above, in the information processing apparatus, learning is performed while an error is propagated in the signal flow direction, and an arithmetic element imitating a nerve cell corresponding to an input / output relationship such as a multiple-to-multiple can be constructed.

  An information processing apparatus according to the present invention is an information processing apparatus including an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells. A plurality of switches that switch an input signal to the arithmetic element between an external input signal from the external input terminal and an output signal from the arithmetic element, and the independent switching unit in which the plurality of switches operate independently; An error calculation unit that calculates an error based on the input signal and the output signal from the calculation element, and the error calculated by the error calculation unit is propagated along the propagation direction of the signal from the output signal of the calculation element to the input signal. And a backward learning unit that adjusts the arithmetic element so as to reduce the error.

  In this information processing apparatus, the external input signal from the external input terminal and the output signal from the computing element are independently switched by each switch in the independent switching unit, and one of the external input signal and the output signal of the computing element is switched. Is input as an input signal of the arithmetic element. In particular, in the information processing apparatus, when learning is performed, when all signals having a relationship to be learned are set to the external input terminal, each signal set to the external input terminal by the independent switching unit is used as an arithmetic element. Each is input as an input signal. In the information processing apparatus, the error calculation unit calculates an error between the input signal and the output signal with respect to the arithmetic element. Further, in the information processing apparatus, an error is propagated in the direction in which the signal flows from the output signal to the input signal by the backward learning unit, and the arithmetic element is adjusted so that the error is reduced. As described above, in the information processing apparatus, learning is performed while propagating an error in a direction opposite to the direction in which the signal flows, and an arithmetic element imitating a neuron corresponding to an input / output relationship such as a multiple-to-multiple can be constructed. it can.

  An information processing apparatus according to the present invention is an information processing apparatus including an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells. A plurality of switches that switch an input signal to the arithmetic element between an external input signal from the external input terminal and an output signal from the arithmetic element, and the independent switching unit in which the plurality of switches operate independently; An error calculation unit that calculates an error based on the input signal and the output signal from the calculation element, and the error calculated by the error calculation unit is propagated along the propagation direction of the signal from the input signal to the output signal of the calculation element. A forward learning unit that adjusts the arithmetic element so as to reduce the error, and an error calculated by the error calculation unit is propagated along the propagation direction of the signal from the output signal of the arithmetic element to the input signal. Characterized in that it comprises a rear learning unit for adjusting the operation device such that the error becomes smaller.

  In this information processing apparatus, in the same manner as described above, when learning is performed, when all signals having a relationship to be learned are set to the external input terminals, each signal set to the external input terminal by the independent switching unit. Are input to the arithmetic elements as input signals. In the information processing apparatus, the error calculation unit calculates an error between the input signal and the output signal with respect to the arithmetic element. Further, in the information processing apparatus, the forward learning unit propagates the error along the signal flowing direction to adjust the arithmetic element, or the backward learning unit propagates the error along the direction opposite to the signal flowing direction. Adjust the computing elements. As described above, in the information processing apparatus, learning is performed while propagating an error in the forward direction and / or the backward direction, and an arithmetic element imitating a nerve cell corresponding to an input / output relationship such as a multiple-to-multiple can be constructed. .

  An information processing method according to the present invention is an information processing method that receives an external input signal, performs an operation simulating a nerve cell by a plurality of arithmetic elements, and outputs an external output signal to the outside. The input signal is switched between an external input signal and an output signal from the arithmetic element, and an independent switching step for independently performing the plurality of switching is included.

  In this information processing method, in the independent switching step, the external input signal and the output signal (feedback signal) from the arithmetic element are independently switched, and one of the external input signal and the output signal of the arithmetic element is switched to the arithmetic element. Input as an input signal. With this information processing method, it is possible to learn a relationship in which a plurality of external output signals correspond to a plurality of external input signals. A plurality of external output signals can be obtained with respect to the input signal.

  An information processing method according to the present invention is an information processing method that receives an external input signal, performs an operation simulating a nerve cell by a plurality of arithmetic elements, and outputs an external output signal to the outside. Based on the independent switching step in which the input signal is switched between the external input signal and the output signal from the arithmetic element, and the switching is independently performed, and the input signal to the arithmetic element and the output signal from the arithmetic element It includes an error calculation step for calculating an error, and a learning step for adjusting an arithmetic element so that the error calculated in the error calculation step is reduced.

  In this information processing method, the external input signal and the output signal from the arithmetic element are independently switched in the independent switching step, and one of the external input signal and the output signal of the arithmetic element is input as the input signal of the arithmetic element. Let In particular, in the information processing method, when learning is performed, all signals input from the external input terminal and having a relationship to be learned are input to the arithmetic elements as input signals in the independent switching step. In the information processing method, the error between the input signal and the output signal for the arithmetic element is calculated in the error calculation step. Further, in the information processing method, the arithmetic element is adjusted so that the error is reduced in the learning step. By this information processing method, it is possible to construct an arithmetic element that simulates a nerve cell corresponding to a plurality of input / output relations.

  A recording medium according to the present invention is a computer-readable recording medium that records an information processing program for receiving an external input signal, performing an operation simulating a nerve cell by a plurality of arithmetic elements, and outputting an external output signal. An information processing program for switching an input signal to a plurality of arithmetic elements between an external input signal and an output signal from the arithmetic element, and causing the computer to execute an independent switching step for performing the switching independently. It is characterized by recording. In addition, the recording medium according to the present invention receives an external input signal, performs an operation simulating a nerve cell by a plurality of arithmetic elements, and records an information processing program for outputting an external output signal. An independent switching step for switching input signals to a plurality of arithmetic elements between an external input signal and an output signal from the arithmetic element, and independently performing the switching, and an input signal to the arithmetic element. An information processing program for causing a computer to execute an error calculation step for calculating an error based on an output signal from an arithmetic element and a learning step for adjusting the arithmetic element so that the error calculated in the error calculation step is reduced It is characterized by recording. According to these recording media, the information processing program recorded on the recording medium is executed on the computer, so that the same effects as those of the information processing methods described above can be obtained.

  It should be noted that the fact that a plurality of input / output relationships can be handled, of course, can correspond to a one-to-one, one-to-multiple, and multiple-to-one input / output relationships.

  ADVANTAGE OF THE INVENTION According to this invention, while being able to learn the multiple-to-multiple input / output relationship in the calculation imitating a neural network, a several output signal can be taken out from several input signals.

  Embodiments of an information processing apparatus, an information processing method, and a recording medium according to the present invention will be described below with reference to the drawings.

  In the present embodiment, the present invention is applied to an information processing apparatus that functions as a neural network simulation apparatus. The information processing apparatus according to the present embodiment is configured by a computer such as a personal computer, and each function operates by executing an information processing program that realizes each function stored in the memory of the computer. In this embodiment, there are three forms. The first embodiment is a basic configuration capable of dealing with the relationship between a plurality of input / output signals, and the second embodiment is a basic structure. The configuration includes an input control function and an operation state determination function for an external input signal in the configuration, and the third embodiment is a configuration at the time of learning for dealing with the relationship between a plurality of input / output signals.

  An information processing apparatus 1 according to the first embodiment will be described with reference to FIG. FIG. 1 is a configuration diagram of an information processing apparatus according to the first embodiment.

  The information processing apparatus 1 is an apparatus that can perform an operation simulating a neural network composed of a large number of nerve cells, and is particularly capable of dealing with a plurality of input / output relationships. The information processing apparatus 1 includes an external input terminal 2, a switching signal input terminal 3, an external output terminal 4, an independent switching unit 5, an input terminal 6, an output terminal 7, and a calculation unit 8.

  The external input terminal 2 is a terminal for inputting an external input signal. The external input terminal 2 has a plurality of terminals 2a,... For inputting a plurality of external input signals, and is connected to each change-over switch of the independent switching unit 5. In the example of FIG. 1, the external input terminal 2 has three terminals 2a, 2b, and 2c. The switching signal input terminal 3 is a terminal for inputting a switching signal for independently switching the changeover switches 5a of the independent switching unit 5. The switching signal input terminal 3 has a plurality of terminals 3a,... According to the number of switching switches of the independent switching unit 5, and is connected to each switching switch of the independent switching unit 5. In the example of FIG. 1, the switching signal input terminal 3 has three terminals 3a, 3b, and 3c. The external output terminal 4 is a terminal for outputting an external output signal to the outside. The external output terminal 4 has a plurality of terminals 4 a... According to the number of unit operators of the calculation unit 8 and is connected to each terminal of the output terminal 7. In the example of FIG. 1, the external output terminal 4 includes six terminals 4a, 4b, 4c, 4d, 4e, and 4f. The output of some operators may not be output externally.

  The independent switching unit 5 switches each external input signal input from the external input terminal 2 and each output signal of the calculation unit 8 independently, and outputs the switched signals to the input terminal 6 of the calculation unit 8. . For this purpose, the independent switching unit 5 includes changeover switches 5 a according to the number of terminals of the external input terminal 2. In the example of FIG. 1, there are three change-over switches 5a, 5b, and 5c. Each change-over switch 5a is connected to a corresponding terminal of the external input terminal 2 and a corresponding terminal of the output terminal 7 at two input ends, respectively, and a corresponding terminal of the input terminal 6 is connected to one output end. The Further, each changeover switch 5a... Receives a changeover signal from a corresponding terminal of the changeover signal input terminal 3, and operates the switch in accordance with the changeover signal. In each change-over switch 5a, one of the external input signal input from the corresponding terminal of the external input terminal 2 and the output signal from the corresponding unit arithmetic element of the arithmetic unit 8 is supplied to the arithmetic unit. 8 is an input signal. The switching signal is a signal indicating whether it is an external input or a feedback of an output signal.

The input terminal 6 is a terminal for sending an input signal to all unit arithmetic elements of the arithmetic unit 8. The input terminal 6 has a plurality of terminals 6a... According to the number of unit operators of the calculation unit 8. In the example of FIG. 1, the input terminal 6 includes six terminals 6a, 6b, 6c, 6d, 6e, and 6f. The output terminal 7 is a terminal for taking out an output signal from each unit arithmetic element of the arithmetic unit 8. The output terminal 7 has a plurality of terminals 7a... According to the number of unit operators of the calculation unit 8. In the example of FIG. 1, the output terminal 7 has six terminals 7 a, 7 b, 7 c, 7 d, 7 e, and 7 f. The calculation unit 8 performs a calculation imitating a neural network composed of a large number of nerve cells. Therefore, the calculation unit 8 includes a large number of unit calculation elements 8a,. In the example of FIG. 1, the calculation unit 8 includes six unit calculation elements 8a, 8b, 8c, 8d, 8e, and 8f. Each unit arithmetic element 8a is constructed so as to be able to cope with a plurality of mutually related relationships by learning. Each unit arithmetic element 8a... Receives input signals from all terminals of the input terminal 6 and outputs output signals to the corresponding terminals of the output terminal 7, respectively. The learning will be described in detail later.

  An output signal is fed back via the independent switching unit 5 to some of the unit arithmetic elements 8a... (A number corresponding to the number of terminals of the external input terminal 2). For this purpose, each terminal of the output terminal 7 corresponding to each part of the unit arithmetic elements is connected to each change-over switch 5a. In the example of FIG. 1, output signals are fed back to the unit arithmetic elements 8 a, 8 b and 8 c via the independent switching unit 5. On the other hand, the other unit arithmetic elements of the unit arithmetic elements 8a... Directly feed back the output signal. Therefore, each terminal of the output terminal 7 corresponding to each other unit arithmetic element is directly connected to a corresponding terminal of the input terminal 6. The reason why the output signal of the unit arithmetic element is always fed back in this way is to set the internal state. In the example of FIG. 1, the output signals of the unit arithmetic elements 8d, 8e, 8f are directly fed back.

  In the calculation unit 8, when an external input signal is input as an input signal, each unit calculation element 8 a... Performs a calculation that simulates a nerve cell. And in the calculating part 8, the output signal of each unit calculating element 8a ... is repeatedly fed back as an input signal, and the calculation imitating a nerve cell is repeatedly performed. When the mutual relation between the signals including the input external input signal is learned, the calculation unit 8 converges the output signal to a certain value while repeatedly executing the calculation, and the external input signal Outputs the correct output signal. At this time, the difference between the input signal (previous output signal) to the calculation unit 8 and the output signal is 0 or a very small value.

  The operation of the information processing apparatus 1 will be described with reference to FIG. First, an external input signal is input to any one of the external input terminals 2. A switching signal is input to each terminal of the switching signal input terminal 3. Among the switching signals, a signal for external input is set as a switching signal for the switching switch of the independent switching unit 5 to which an external input signal is input, and an output signal is fed back to the switching signals for the other switching switches. Is set. In the independent switching part 5, it switches independently according to a switching signal by each switch 5a .... At this time, the changeover switch to which the changeover signal for which the signal for external input is set is inputted switches to the external input signal as the input signal and outputs to the input terminal 6. On the other hand, in the change-over switch to which a change-over signal in which a signal for feedback is set is input, the change-over switch is switched to the output signal of the unit arithmetic element as an input signal and output to the input terminal 6.

And each unit arithmetic element 8a ... of the calculating part 8 repeatedly performs the calculation which imitated the nerve cell, and outputs an output signal, respectively. For each calculation, the output signal of each unit calculation element 8a... Of the calculation unit 8 is fed back to all unit calculation elements as an input signal via the independent switching unit 5 or directly. Here, an input signal at time t of each unit arithmetic element of the arithmetic unit 8 is assumed to be x _j (t). j is a number assigned to each terminal of the input terminal 6, and is a number (1, 2,...) from 1 to the number of terminals of the input terminal 6. The o _i output of the i-th unit operation element is represented by the formula (1). i is a number assigned to each unit computing element of the computing unit 8, and is a number (1, 2,...) from 1 to the number of unit computing elements of the computing unit 8.

ｆ（）は単位演算素子における非線形演算を行うための非線形関数であり、Σは単位演算素子における重みを付けられた全ての入力信号の加算演算を表し、ｗ_ｉｊは入力端子６のｊ番目の端子とｉ番目の単位演算素子との間のシナプス荷重器の荷重値である。さらに、演算部８の各単位演算素子の出力信号の時間変化は、式（２）で表される。

f () is a non-linear function for performing a non-linear operation in the unit arithmetic element, Σ represents an addition operation of all weighted input signals in the unit arithmetic element, and w _ij is the j th of the input terminal 6. It is a load value of the synapse loader between the terminal and the i-th unit arithmetic element. Furthermore, the time change of the output signal of each unit arithmetic element of the arithmetic unit 8 is expressed by Expression (2).

この式（２）から、演算部８の各単位演算素子の入力信号と出力信号とが等しくなるかあるいはほぼ等しくなるかを判断できる。情報処理装置１では、この演算部８における演算が繰り返し実行されることによって各単位演算素子の入力信号と出力信号とが等しくなったとき安定状態となる。このとき、情報処理装置１では、演算部８では入力された外部入力信号に対してつじつまの合った出力信号を出力しており、この出力信号を外部出力信号として外部出力端子４から出力する。この際、内部状態の設定に応じて出力される外部出力信号が変わる。つまり、任意の外部入力信号に対して、内部状態に応じて、つじつまの合った複数の外部出力信号を出力することができる。

From this equation (2), it can be determined whether the input signal and the output signal of each unit arithmetic element of the arithmetic unit 8 are equal or nearly equal. The information processing apparatus 1 is in a stable state when the input signal and the output signal of each unit calculation element are equalized by repeatedly executing the calculation in the calculation unit 8. At this time, in the information processing apparatus 1, the arithmetic unit 8 outputs an output signal that is consistent with the input external input signal, and outputs this output signal from the external output terminal 4 as an external output signal. At this time, the external output signal to be output changes according to the setting of the internal state. That is, it is possible to output a plurality of appropriate external output signals for any external input signal according to the internal state.

  Further, when another external input signal is input from a terminal other than the terminal first input to the external input terminal 2, the information processing apparatus 1 also uses the newly input external input signal as described above. On the other hand, a suitable external output signal is output from the external output terminal 4. In this way, the information processing apparatus 1 can obtain a plurality of external output signals that are consistent with any external input signal among the plurality of external input signals.

  Here, as a relationship between a plurality of pairs, a word will be described as an example. For example, the name (cedar, pine, beech) is associated with the concept of (wood, wood). Specifically, the associations of cedar → wood, pine → wood, beech → wood, and the associations of cedar → wood, pine → wood, and beech → wood are established. On the other hand, the association of wood → cedar, wood → pine, wood → beech, and the association of wood → cedar, wood → pine, and wood → beech are also established. Therefore, there is a multiple-to-multiple relationship between the components of (cedar, pine, beech) and (wood, wood). When learning such a mutual relationship between words, the information processing apparatus 1 outputs “cedar”, “pine”, or “beech” according to the internal state when “tree” is input. , "Wood" will output either "cedar", "pine", or "beech" depending on the internal state, and conversely, any of "cedar", "cedar", or "beech" When input, either “wood” or “wood” is output according to the internal state.

  In addition, this multiple-to-multiple relationship of words creates some interpretation limitations in the context of topics in human communication, and is associated with an appropriate correspondence. For example, in the world heritage topic, when the Shirakami Mountains, where the beech primeval forest is famous, is the topic, the word “beech” is strongly related to the name “tree”, not the name “wood”. It is appropriate to be attached. When such a thing is also learned, the information processing apparatus 1 can output an appropriate correspondence according to the topic. In this way, in the information processing apparatus 1, when learning is performed in advance, it is possible to extract a relationship corresponding to a plural-to-multiple relationship of words from a conversation with a person, and human semantic interpretation is possible. It is also possible to realize an interface having a conversation function.

  The information processing apparatus 1 basically outputs an external output signal that is consistent with the external input signal. However, when the external output signal is not consistent with the external input signal (that is, The external output signal may be output even when there is a contradiction between the external input signal and the external output signal. At this time, in order to clearly show the contradictory portion, when the information processing apparatus 1 is provided with a monitor, the contradictory portion is displayed by color, a change in contrast, a change in texture, a font Or by changing the sound if an audio output device is provided. Even in the case where there is no mutual relationship between the external input signal and the external output signal, there is a possibility that a neural network that performs a new idea that cannot be considered by common sense in the arithmetic unit 8 may be constructed. Because there is, it is also important to output the contradiction. This is particularly important when the purpose of information processing is to support ideas. An example of the contradiction is a contradiction in which the name of a bird such as “crow” is output when “tree” is input, for example.

  According to the information processing apparatus 1, by providing a function of independently switching the input to the calculation unit 8 between an external input signal and a feedback signal (output signal of the calculation unit 8), a plurality of pairs (of course, one-to-multiple). It is possible to take out an external output signal that is consistent with an arbitrary external input signal among a plurality of external input signals in a mutual relationship (including a multiple-to-one relationship and a one-to-one relationship). At this time, a plurality of external output signals can be extracted according to the internal state. Further, according to the information processing apparatus 1, even when there is a contradiction between the external input signal and the external output signal, the contradictory information can be obtained.

  Furthermore, by adding a learning function to the information processing apparatus 1, it is possible to learn the mutual relationship between a plurality of signals and to simulate a neural network having a plurality of relationships. Part 8 can be constructed.

  With reference to FIG. 2, an information processing apparatus 11 according to the second embodiment will be described. FIG. 2 is a configuration diagram of an information processing apparatus according to the second embodiment. In the information processing apparatus 11, the same reference numerals are given to the same configurations as those of the information processing apparatus 1 according to the first embodiment, and the description thereof is omitted.

  The information processing apparatus 11 is the same apparatus as the information processing apparatus 1 and further includes an input control function and a calculation state determination function for an external input signal. For this purpose, the information processing apparatus 11 includes an external input terminal 2, a switching signal input terminal 3, an external output terminal 4, an independent switching unit 5, an input terminal 6, an output terminal 7, an arithmetic unit 8, an external input control unit 12, A definite output unit 13 and a calculation state determination unit 14 are provided.

  The external input control unit 12 controls the input order of a plurality of input external input signals, and causes the independent switching unit 5 to sequentially input the external input signals according to a predetermined order. The external input control unit 12 is connected to each terminal 2a of the external input terminal 2 on the input side, and connected to the changeover switch 5a of the independent switching unit 5 on the output side. Further, the calculation state determination unit 14 is connected to the external input control unit 12 and each output confirmation signal is input. When a plurality of external input signals are input to the external input terminal 2, the external input control unit 12 sets the input order. This input order is defined in advance according to the input signal. The external input control unit 12 first outputs the first external input signal in the input order to the independent switching unit 5. When the output confirmation signal is switched from output unconfirmed to output confirmation, the external input control unit 12 stops the output of the external input signal that has been output and triggers the external input signal next in the input order as the independent switching unit 5. Output to. The external input control unit 12 repeatedly performs such processing, and sequentially inputs a plurality of external input signals according to the input order.

  The importance of controlling the input order of a plurality of external input signals in this way will be described. One of the importance will be described. In the information processing apparatus 11 including the arithmetic unit 8 that has learned two mutually related signals A and B, a signal A is output as an external input signal and a signal B is output as an external output signal. There is a switch between when A is output as an external output signal. This corresponds to switching between one reasoning and vice versa. In general, the opposite is not true even if some reasoning is correct. However, humans have the characteristic that they can make inferences that include such theoretically incorrect inferences, leading to various ideas. In this way, in order to realize information processing close to human beings that can extract the relationship between signals even if the reasoning is incorrect from a logical point of view, the input order of multiple external input signals It is necessary to control.

  The other importance will be explained. This is because the multiple-to-multiple relationship is controlled by the context. By controlling the order of external input signals (words constituting a context), information can be narrowed down from a context having an arbitrary order. This means that even when the same plurality of external input signals (words) are given, different conclusions can be reached due to the difference in the input order. In other words, even when the same plurality of words are used, if the context is generated by changing the order of the words, the meaning may change. For example, when a set of four words “young”, “male”, “female”, and “to” is given, “young man and woman” and “man and young woman” have different meanings. The former means either “(young man) and (woman)” or “young (man and woman)”. On the other hand, the latter has one meaning of “(male) and (young woman)”. Thus, even the same word set may have different meanings depending on the order. In order to eliminate the diversity of meanings of such word sets, it is necessary to converge to the intended one (or very few) meanings by appropriately selecting the order of the words. Therefore, it is necessary to control the input order of a plurality of external input signals.

  The confirmed output unit 13 outputs, to the external output terminal 4, an output signal that has been determined by the calculation state determination unit 14 that the calculation state has been fixed among the output signals output from the calculation unit 8. For this purpose, the deterministic output unit 13 has an output switch 13a... For each unit arithmetic element 8a. In the example of FIG. 2, there are six output switches 13a, 13b, 13c, 13d, 13e, and 13f. Each output switch 13a... Has one input terminal connected to the corresponding terminal of the output terminal 7, and one output terminal connected to the corresponding terminal of the external output terminal 4. Further, each output switch 13a... Receives an output confirmation signal from the calculation state determination unit 14 and operates the switch in accordance with the output confirmation signal. At this time, each output switch 13a... Is turned off when the output confirmation signal is a signal indicating that the output has not been confirmed, and is turned on when the output confirmation signal is a signal indicating that the output has been confirmed. Output as an external output signal.

  The calculation state determination unit 14 determines the calculation state of the calculation unit 8 and determines whether or not the output signal has converged to a certain value. Therefore, the calculation state determination unit 14 takes in the output signal and input signal (that is, the previous output signal) of each unit calculation element 8a. Then, the calculation state determination unit 14 substitutes the output signal and the input signal for a predetermined state determination function. The predetermined state determination function is a function for detecting a time change of the output signal of each unit arithmetic element 8a... Of the arithmetic unit 8, and becomes 0 when the difference between the output signal and the input signal becomes 0. For example, the function shown in Expression (2). The calculation state determination unit 14 sets the output determination to the output determination signal when the calculation result of the state determination function is 0 or below a certain value, and sets the output unconfirmed otherwise. Note that the output confirmation signal is a signal indicating whether or not the output signal of each unit arithmetic element 8a... Of the arithmetic unit 8 has been determined to a certain value. A signal indicating that the output is not determined is set when the signal indicating is set and fluctuating.

  When there is no mutual relationship between the signals, the information processing apparatus does not become stable or the error between the input signal and the output signal of the arithmetic unit 8 does not become a certain value or less, so there is a mutual relationship between the signals. It can be determined whether or not. Therefore, it is possible to determine whether or not there is a mutual relationship between the signals by determining whether or not the output signal is determined by the calculation state determination unit 14.

  The information processing apparatus 11 has the same effects as the information processing apparatus 1 according to the first embodiment, and also has the following effects. The information processing apparatus 11 can extract the relationship between different signals by controlling the input order of a plurality of external input signals. Therefore, it is possible to realize an information processing similar to a human and an interface having a conversation function that enables human semantic interpretation. In addition, since the information processing apparatus 11 is configured to output to the outside after the output signal in the arithmetic unit 8 settles to a certain value, only the determined value can be taken out outside, and there is a mutual relationship between the signals. Only certain values can be retrieved.

  With reference to FIG. 3, an information processing apparatus 21 according to a third embodiment will be described. FIG. 3 is a configuration diagram of an information processing apparatus according to the third embodiment. Note that, in the information processing device 21, the same components as those of the information processing device 1 according to the first embodiment are denoted by the same reference numerals, and description thereof is omitted. In FIG. 3, each terminal of the external input terminal 2, each terminal of the switching signal input terminal 3, each terminal of the external output terminal 4, each switch of the independent switching unit 5, each terminal of the input terminal 6, and each of the output terminal 7 Although only two terminals are drawn, there are actually many terminals. In FIG. 3, only the first unit arithmetic element 8a is illustrated in detail, but the other unit arithmetic elements have the same configuration, and each unit arithmetic element has the same input from the input terminal 6. A signal is input and an output signal is output to the corresponding terminal of the output terminal 7. In the information processing apparatus 21, the same learning is performed for each unit arithmetic element 8a.

  The information processing device 21 is a device for learning a mutual relationship between a plurality of signals and adjusts the calculation unit 8 (particularly, the load value of the synapse loader of each unit calculation element). For this purpose, the information processing apparatus 21 includes an external input terminal 2, a switching signal input terminal 3, an external output terminal 4, an independent switching unit 5, an input terminal 6, an output terminal 7, a calculation unit 8, an error calculation unit 25, and a backward learning unit. 26, a forward learning unit 27 is provided. Incidentally, the arithmetic unit 8 adjusted by learning can output a plurality of output signals consistent with each external input signal.

Each unit arithmetic element 8a... Of the arithmetic unit 8 includes a synapse loader 22a..., An adder 23, and a nonlinear arithmetic unit 24. The synapse loader 22a... Is a loader that imitates the synapse connection of nerve cells and has a load value w _ij . Each synapse loader 22a... Receives a corresponding input signal from each terminal of the input terminal 6, multiplies the input signal by a load value w _ij , and outputs it to the adder 23. The adder 23 is an adder that calculates the sum of the input signals weighted by the synapse loaders 22a. The nonlinear computing unit 24 is a computing unit that performs nonlinear computation on the total value of the adder 23, and outputs the computation result as an output signal to a corresponding terminal of the output terminal 7. Examples of the nonlinear function of the nonlinear computing unit 24 include a sigmoid function and a snake side function. The output o _i of the i-th unit arithmetic element when the input signal to the arithmetic unit 8 at time t is x _j (t) is expressed by the above-described equation (1).

  The error calculator 25 calculates an error between the corresponding input signal and output signal of the calculator 8. Therefore, the input signal from each terminal of the input terminal 6 and the output signal from each terminal of the output terminal 7 are input to the error calculation unit 25 for each calculation (every output signal is output) in the calculation unit 8. Is done. Then, the error calculation unit 25 calculates the difference between the corresponding input signal and output signal, and outputs the difference to the backward learning unit 26 and the forward learning unit 27 as an error signal.

  The backward learning unit 26 propagates the error signal along the propagation direction of the signal from the output signal of the calculation unit 8 to the input signal (that is, the direction opposite to the signal flow direction), so that the error is reduced. 8 (especially, the load value of each synapse loader) is adjusted. The forward learning unit 27 propagates the error signal along the signal propagation direction (that is, the signal flow direction) from the input signal to the output signal of the calculation unit 8, and the calculation unit 8 (particularly, Adjust the load value of each synapse loader. Therefore, each time an error is calculated by the error calculation unit 25, the error calculated by the error calculation unit 25, the load value of each synapse loader 22 a. The calculation result (that is, the output signal) of the calculator 24 is input, and the adjusted load value is output to each synapse loader 22a.

With reference to FIG. 3, the operation of the information processing apparatus 21 when learning the relationship between mutually related signals will be described. First, as an initial state, in the information processing apparatus 21, all the output signals of the arithmetic unit 8 are switched to be sent to the respective terminals of the input terminal 6 by the respective switches 5 a of the independent switching unit 5. At this time, the information processing device 21 outputs toward the state determined by each load value of the synapse loader 22a... Of the arithmetic unit 8 simulating the neural network and the nonlinear function of the nonlinear operator 24 according to the principle of feedback. The signal o _i (t) is changed.

Next, signals related to each other to be learned are set in each terminal of the external input terminal 2. Then, in the information processing apparatus 21, switching is performed so that all signals set to the external input terminal 2 are sent to the corresponding terminals of the input terminal 6 by the respective switches 5a. The external input signal is input to the calculation unit 8 from each terminal. For example, if the signals related to each other are x ₁ ^E (t) and x ₂ ^E (t), they are set to the first and second terminals of the external input terminal 2, respectively, and this x ₁ ^E (t) and x ₂ ^E (t) is input to the terminals 6 a and 6 b of the input terminal 6 through the change-over switches 5 a and 5 b and further input to the calculation unit 8. In order to make the explanation easy to understand, _two signals x ₁ ^E (t) and x ₂ ^E (t) are set as interrelated signals to be learned. A signal may be set.

Then, in the information processing device 21, each load of the synapse loaders 22 a... Of the calculation unit 8 is determined based on each load value of the synapse loaders 22 a. In addition to the value and the nonlinear function of the nonlinear computing unit 24, the output signal of the computing unit 8 is changed toward the state where the inputs of the terminals 6a and 6b of the input terminal 6 are x ₁ ^E (t) and x ₂ ^E (t). Let

When the information processing apparatus 21 has already learned the mutual relationship between x ₁ ^E (t) and x ₂ ^E (t), these two signals x ₁ ^E (t) and x ₂ ^E (t) Are simultaneously output as external output signals o ₁ ^E (t) and o ₂ ^E (t). On the other hand, when learning is not performed, the relationship is expressed by the equations (3) and (4).

式（３）、（４）の関係にある場合、演算部８の対応する入力信号と出力信号との間には差があり、これが誤差となる。誤差計算部２５では、演算部８から出力信号ｏ_ｉ（ｔ）が出力される毎に、演算部８の対応する入力信号ｘ_ｉ（ｔ）と出力信号ｏ_ｉ（ｔ）とから式（５）により誤差ｅ_ｉ（ｔ）を計算する。例えば、演算部８の入力信号ｘ_１（ｔ）と出力信号ｏ_１（ｔ）の差から誤差ｅ_１（ｔ）が計算され、入力信号ｘ_２（ｔ）と出力信号ｏ_２（ｔ）とから誤差ｅ_２（ｔ）が計算される。

In the case of the relationship of Expressions (3) and (4), there is a difference between the corresponding input signal and output signal of the calculation unit 8, and this is an error. Each time the output signal o _i (t) is output from the calculation unit 8, the error calculation unit 25 calculates the equation (5) from the corresponding input signal x _i (t) and output signal o _i (t) of the calculation unit 8. ) To calculate the error e _i (t). For example, the error e ₁ (t) is calculated from the difference between the input signal x ₁ (t) and the output signal o ₁ (t) of the calculation unit 8, and the input signal x ₂ (t) and the output signal o ₂ (t) An error e ₂ (t) is calculated from

そして、情報処理装置２１では、後方学習部２６又は／及び前方学習部２７により、この計算した各誤差が小さくなるようにシナプス荷重器２２ａ・・・の各荷重値を変化させ、学習を行う。この際、そして、後方学習部２６又は／及び前方学習部２７では、誤差計算部２５で誤差ｅ_ｉ（ｔ）が計算される毎に、誤差ｅ_ｉ（ｔ）が小さくなるように荷重値ｗ_ｉｊを変化させる。そして、演算部８の各単位演算素子８ａ・・・では、変化された荷重値ｗ_ｉｊを用いて、演算を行い、出力信号を出力する。この一連の処理が繰り返し行われ、誤差ｅ_ｉ（ｔ）が小さくなっていき、荷重値ｗ_ｉｊが調整されていく。やがて、荷重値ｗ_ｉｊが最適な値に調整され、外部入力端子２に設定された外部入力信号が外部出力信号として出力されるようになる。例えば、外部入力信号ｘ_１ ^Ｅ（ｔ），ｘ_２ ^Ｅ（ｔ）が外部出力信号ｏ_１ ^Ｅ（ｔ），ｏ_２ ^Ｅ（ｔ）として出力されるようになる。以下で、後方学習部２６と前方学習部２７による学習の方法についてそれぞれ説明する。

Then, in the information processing apparatus 21, the backward learning unit 26 and / or the forward learning unit 27 performs learning by changing each load value of the synapse loaders 22a... So that each calculated error becomes small. At this time, in the backward learning unit 26 and / or the forward learning unit 27, each time the error e _i (t) is calculated by the error calculation unit 25, the load value w so that the error e _i (t) is reduced. Change _ij . And each unit arithmetic element 8a ... of the calculating part 8 calculates using the changed load value _wij , and outputs an output signal. This series of processing is repeated, and the error e _i (t) is reduced, and the load value w _ij is adjusted. Eventually, the load value w _ij is adjusted to an optimum value, and the external input signal set in the external input terminal 2 is output as the external output signal. For example, the external input signals x ₁ ^E (t) and x ₂ ^E (t) are output as the external output signals o ₁ ^E (t) and o ₂ ^E (t). Hereinafter, learning methods by the backward learning unit 26 and the forward learning unit 27 will be described.

First, a learning method by the backward learning unit 26 will be described. Similar to the error back-propagation method, the relationship between necessary signals in the stable state is obtained, and therefore the backward learning unit 26 uses the error e _i (t) obtained by Equation (5) to express the cost function. In (6), each load value w _ij of the synapse loader 22a... Is changed so that cost becomes small. In this case, the change amount _{dw ij} of each weight value _{w ij} is given by Equation (7). In the equation (7), γ is a positive constant. Here, de _a / dw _ij is given in the form of equation (8) by proceeding with the calculation according to the derivation of the error back propagation method.

式（８）の右辺のＫ_ｉｊは、演算子の間に接続がある場合には１、接続がない場合には０である構造因子である。通常、構造因子は、初期状態で定めたままで固定するが、必要に応じて変化させてもよい。また、（Ｉ−Ｆ’Ｗ）の逆行列（Ｉ−Ｆ’Ｗ）^−１は、これを正確に計算してもよいが、（Ｉ−Ｆ’Ｗ）^−１＝１＋Ｆ’Ｗ＋（Ｆ’Ｗ）^２＋（Ｆ’Ｗ）^３＋（Ｆ’Ｗ）^４＋・・・と形式的に展開したものの低次の項の和をもって近似してもよい。

K _{ij on} the right side of Equation (8) is a structural factor that is 1 when there is a connection between operators and 0 when there is no connection. Usually, the structural factor is fixed as it is in the initial state, but may be changed as necessary. In addition, the inverse matrix (IF′W) ^{−1 of} (IF−W) may calculate this accurately, but (IF−W) ⁻¹ = 1 + F′W + (F ′ W) ² + (F′W) ³ + (F′W) ⁴ +... May be approximated by the sum of the low-order terms.

Next, a learning method by the forward learning unit 27 will be described. The forward learning unit 27 changes the load value w _ij of the synapse loader 22a... More directly by paying attention to the internal state, thereby changing the error e _i (t) obtained by Expression (5). Make it smaller. That is, the load value change amount dw _ij that satisfies the relationship of Expression (9) is obtained, and the load value w _ij is changed accordingly. The change amount dw _ij of the load value at this time is the value shown in the equation (11) because the value of the primary order may be the relationship of the equation (10).

式（１０）のλは学習速度を決めるパラメータである。一次オーダーの荷重値ｗ_ｉｊの変化を代入して計算を進めることにより、二次以上のオーダーの項も計算できる。このように、この学習方法は、誤差を信号の流れる方向に伝播させることに対応するので、前方学習と呼んでいる。

In equation (10), λ is a parameter that determines the learning speed. By substituting the change in the load value w _ij of the primary order and proceeding with the calculation, it is possible to calculate the terms of the secondary order and higher. Thus, this learning method is called forward learning because it corresponds to propagating an error in the direction of signal flow.

  According to the information processing apparatus 21, a learning function is added, and a function for independently switching an input to the calculation unit 8 between an external input signal and a feedback signal (output signal) is provided, so that a plurality of signals can be transmitted between a plurality of signals. You can learn the mutual relationship. By constructing the calculation unit 8 by this learning, it is possible to extract a plurality of external output signals that are consistent with an arbitrary external input signal among a plurality of external input signals in a plurality of pairs. . In particular, the information processing apparatus 21 can perform learning while propagating an error along the direction of signal flow, and can also perform learning while propagating an error along the reverse direction of signal flow.

  As mentioned above, although embodiment which concerns on this invention was described, this invention is implemented in various forms, without being limited to the said embodiment.

  For example, although the present embodiment is configured by a computer in which a program for realizing each function is stored in advance, the program for realizing each function can be recorded on a computer-readable recording medium and distributed. is there. As such a recording medium, for example, a magnetic medium such as a hard disk and a flexible disk, an optical medium such as a CD-ROM and a DVD-ROM, a magneto-optical medium such as a floppy disk, or a program instruction is executed or stored. Specially arranged hardware devices such as RAM, ROM and semiconductor non-volatile memory. Further, a recording medium reading drive (for example, a flexible disk drive) for reading a program or the like from such a recording medium may be connected to the information processing apparatus. It is also possible to distribute a program that realizes each function by downloading it via a network. Also, each unit can be configured by hardware.

  In the third embodiment, the forward learning unit and the backward learning unit are configured. However, even in the configuration including any one of the learning units, a mutual relationship between a plurality of signals is learned, and an arithmetic unit Can be built.

  In this embodiment, the operator is configured by linear transformation and nonlinear transformation by a sigmoid function. However, the configuration of the operator is not limited to these. For example, the function f may be non-monotonous. Alternatively, the linear transformation part may include high-order terms such as the second order and third order of the input, or may be a more complicated function than that.

It is a block diagram of the information processing apparatus which concerns on 1st Embodiment. It is a block diagram of the information processing apparatus which concerns on 2nd Embodiment. It is a block diagram of the information processing apparatus which concerns on 3rd Embodiment.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1,11,21 ... Information processing apparatus, 2 ... External input terminal, 3 ... Switching signal input terminal, 4 ... External output terminal, 5 ... Independent switching part, 5a, 5b, 5c ... Switch, 6 ... Input terminal, 7 ... output terminal, 8 ... arithmetic unit, 8a, 8b, 8c, 8d, 8e, 8f ... unit arithmetic element, 12 ... external input control unit, 13 ... confirmed output unit, 13a, 13b, 13c, 13d, 13e, 13f ... Output switch, 14... Arithmetic state determination unit, 22 a, 22 b, 22 c, 22 d, 22 e, 22 f... Synaptic loader, 23 ... adder, 24 ... nonlinear operation unit, 25 ... error calculation unit, 26 ... backward learning unit, 27 … Forward learning department

Claims (12)

An information processing apparatus comprising an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells,
An independent switching unit having a plurality of switches for switching input signals to the plurality of arithmetic elements between an external input signal from the external input terminal and an output signal from the arithmetic element, wherein the plurality of switches operate independently; An information processing apparatus comprising:   The information processing apparatus according to claim 1, further comprising an arithmetic state determination unit that determines a processing state in the information processing apparatus based on an input signal to the arithmetic element and an output signal from the arithmetic element.   3. The information processing apparatus according to claim 1, wherein the external output signal is output even when there is a contradiction between the external input signal and the external output signal.   The information processing apparatus according to claim 3, wherein a portion having the contradiction is clearly indicated.   5. The information processing apparatus according to claim 1, further comprising an external input control unit that controls an external input signal to be sequentially input to the arithmetic elements. 6. An information processing apparatus comprising an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells,
An error calculator that calculates an error based on an input signal to the arithmetic element and an output signal from the arithmetic element;
A forward learning unit that propagates the error calculated by the error calculation unit along a propagation direction of a signal from the input signal to the output signal of the calculation element and adjusts the calculation element so that the error is reduced. An information processing apparatus characterized by the above. An information processing apparatus comprising an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells,
An independent switching unit that has a plurality of switches that switch an input signal to the plurality of arithmetic elements between an external input signal from the external input terminal and an output signal from the arithmetic element, and the plurality of switches operate independently; ,
An error calculator that calculates an error based on an input signal to the arithmetic element and an output signal from the arithmetic element;
A backward learning unit for propagating the error calculated by the error calculation unit along the propagation direction of the signal from the output signal of the calculation element to the input signal and adjusting the calculation element so that the error is reduced; An information processing apparatus characterized by the above. An information processing apparatus comprising an external input terminal that receives an external input signal, an external output terminal that outputs an external output signal to the outside, and a plurality of arithmetic elements that simulate nerve cells,
An independent switching unit that has a plurality of switches that switch an input signal to the plurality of arithmetic elements between an external input signal from the external input terminal and an output signal from the arithmetic element, and the plurality of switches operate independently; ,
An error calculator that calculates an error based on an input signal to the arithmetic element and an output signal from the arithmetic element;
A forward learning unit that propagates an error calculated by the error calculation unit along a propagation direction of a signal from an input signal of the calculation element to an output signal, and adjusts the calculation element so as to reduce the error;
A backward learning unit for propagating the error calculated by the error calculation unit along the propagation direction of the signal from the output signal of the calculation element to the input signal and adjusting the calculation element so that the error is reduced; An information processing apparatus characterized by the above. An information processing method for receiving an external input signal, performing an operation simulating a nerve cell by a plurality of arithmetic elements, and outputting an external output signal,
An information processing method comprising: an independent switching step of switching input signals to the plurality of arithmetic elements between an external input signal and an output signal from the arithmetic elements, and performing the plurality of switching independently. An information processing method for receiving an external input signal, performing an operation simulating a nerve cell by a plurality of arithmetic elements, and outputting an external output signal,
An independent switching step in which input signals to the plurality of arithmetic elements are respectively switched between an external input signal and an output signal from the arithmetic element, and the plurality of switching is independently performed;
An error calculating step for calculating an error based on an input signal to the arithmetic element and an output signal from the arithmetic element;
And a learning step of adjusting the arithmetic element so that the error calculated in the error calculation step is reduced. A computer-readable recording medium that records an information processing program for receiving an external input signal, performing an operation that simulates a nerve cell by a plurality of arithmetic elements, and outputting an external output signal,
An information processing program for switching an input signal to the plurality of arithmetic elements between an external input signal and an output signal from the arithmetic element and causing the computer to execute an independent switching step for independently performing the plural switching is recorded. Computer-readable recording medium. A computer-readable recording medium that records an information processing program for receiving an external input signal, performing an operation that simulates a nerve cell by a plurality of arithmetic elements, and outputting an external output signal,
An independent switching step in which input signals to the plurality of arithmetic elements are respectively switched between an external input signal and an output signal from the arithmetic element, and the plurality of switching is independently performed;
An error calculating step for calculating an error based on an input signal to the arithmetic element and an output signal from the arithmetic element;
A computer-readable recording medium recording an information processing program for causing a computer to execute a learning step of adjusting the arithmetic element so that an error calculated in the error calculation step is reduced.

Images