JP2007220032A - Method and device for recognizing time-series data, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a time-series data recognizing method for improving recognition precision by reflecting the transition of the time-series data on data recognition, and to provide a time-series data recognizing device and a program. <P>SOLUTION: A winner unit is selected based on similarity between each input vector x(t) constituting the time-series data and a connection weight vector Wj and similarity between a winner afterglow vector Y(t) and an experience weight vector Uj. Then, the winner afterglow vector Y(t) is updated according to the selected winner unit. Besides, the time-series data for recognition is recognized by a part or the whole of the selected winner unit. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a time-series data recognition method, a time-series data recognition apparatus, and a program, and more particularly to recognition of time-series data whose contents change with time.

  A self-organizing map (SOM), which is a kind of neural network, is an unsupervised competitive learning model proposed by T. Kohonen. Self-organization means changing and adapting itself to external signals, and a model that simplifies the self-organization process is a self-organization map. According to such a self-organizing map, it is possible to classify data given from the outside according to its contents, and its application to data recognition technology has been widely studied.

  By the way, the data to be targeted by the data recognition technology includes, for example, voice data that specifies a time-series sound pressure data group, moving image data that specifies a time-series still image group, and handwriting that specifies a time-series writing position group. Most of the data, such as character data, is time-series data whose contents change over time. However, many of the conventional time-series data recognition techniques using a self-organizing map treat the data contents and the time information in the same row and input them to the self-organizing map. For this reason, it is difficult to reflect the transition of time-series data in data recognition, and there is a limit to improving the accuracy of data recognition.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a time-series data recognition method and time-series data recognition capable of improving the recognition accuracy by reflecting the transition of time-series data in data recognition. To provide an apparatus and a program.

  In order to solve the above-described problem, a time-series data recognition method according to the present invention includes a competitive layer in which a plurality of units are arranged, and a self-organization in which first and second weight vectors are associated with each unit. A learning step for learning the learning map using the time series data for learning, and a recognition step for recognizing the time series data for recognition using the self-organizing map.

  The learning step sequentially receives a learning input vector corresponding to the learning time-series data, the sequentially received learning input vector, and the first weight vector associated with each unit, A step of selecting a part of the plurality of units based on the degree of similarity, and a step of bringing the first weight vector associated with the selected unit closer to the accepted learning input vector A step of accepting learning input vectors corresponding to the time-series data for learning in time order, and the similarity between the sequentially accepted learning input vectors and the first weight vectors associated with the units. And selecting a part of the plurality of units based on the selected unit. Among the selected some units, based on the similarity between the second weight vector associated with and the winner reverberation vector indicating the history of the winner unit with respect to the previously accepted learning input vector. Selecting one winner unit from the step, bringing the second weight vector associated with the selected winner unit closer to the winner reverberation vector, and updating the winner reverberation vector according to the selected winner unit Including the steps of:

  On the other hand, the recognition step includes a step of receiving a recognition input vector corresponding to the time-series data for recognition in time order, the recognition input vector sequentially received, and the first weight vector associated with each unit. And selecting a part of the plurality of units based on the similarity, the second weight vector associated with the selected part of the unit, and the recognition received previously. Selecting one winner unit from among the selected units based on the similarity between the winner reverberation vector indicating the history of the winner unit with respect to the input vector, and according to the selected winner unit The step of updating the winner reverberation vector and a part or all of the selected winner units When including the, and step recognizes the series data.

  The time-series data recognition apparatus according to the present invention is a time-series data recognition apparatus for recognizing time-series data using a self-organizing map having a competitive layer in which a plurality of units are arranged. A weight vector storage means for storing the first and second weight vectors, a recognition input vector reception means for receiving the recognition input vectors corresponding to the time-series data for recognition in time order, and the recognition that has been sequentially received. The similarity between the input vector and the first weight vector, and the second weight vector and a winner reverberation vector indicating the history of the winner unit for the recognized input vector previously received by the recognized input vector receiving means. A winner unit selection means for selecting one winner unit from the plurality of units based on the similarity; A winner reverberation vector updating means for updating the winner reverberation vector according to the selected winner unit, and a recognition means for recognizing the time-series data for recognition by a part or all of the selected winner units. It is characterized by.

  A program according to the present invention is a program for recognizing time-series data using a self-organizing map having a competitive layer in which a plurality of units are arranged, and is associated with each unit, Weight vector storage means for storing two weight vectors, recognition input vector reception means for receiving recognition input vectors corresponding to time series data for recognition in time order, the recognition input vectors received sequentially, and the recognition received sequentially The similarity between the input vector and the first weight vector, and the second weight vector and a winner reverberation vector indicating the history of the winner unit for the recognized input vector previously received by the recognized input vector receiving means. Based on the similarity, the winner unit selects one winner unit from the plurality of units. Recognition for recognizing the time-series data for recognition by a selection means, a winner reverberation vector updating means for updating the winner reverberation vector in accordance with the selected winner unit, and a part or all of the selected winner units A program for causing a computer to function as means.

  In the present invention, time series data for recognition is input at the time of data recognition. This time-series data consists of a sequence of recognition input vectors in time order, and each recognition input vector is input in time order. Each unit placed in the competitive layer of the self-organizing map is associated with two weight vectors, namely a first and a second weight vector. The first weight vector is for calculating the similarity (first similarity) with the recognition input vector, and the second weight vector is the similarity (second) with the winner reverberation vector indicating the history of the winner unit. Of similarity) is calculated. According to the present invention, which of the competitive layer units is the winner unit is evaluated based on the first similarity, that is, the contents of the recognition input vector, and the second similarity, that is, the history of the winner unit. And can be determined in consideration of the evaluation. In other words, in the present invention, the winner unit can be determined in consideration of the content of the recognition input vector and the change of the recognition input vector up to the content. For this reason, the transition of the time series data for recognition can be reflected in data recognition, and recognition accuracy can be improved.

  The winner reverberation vector may include a winner reverberation value associated with each unit and updated as the unit is selected as a winner unit. The winner reverberation value may be decreased as the learning input vector or the recognition input vector is accepted.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

  In the time-series data recognition method according to the present embodiment, a self-organizing map is used. FIG. 1 is a diagram showing a network structure of this self-organizing map. In the present embodiment, a self-organizing map having the structure shown in FIG. 3 is managed on a computer, and time series data is recognized by the computer using this map. As shown in the figure, the self-organizing map used in the time-series data recognition method according to the present embodiment is characterized by being composed of three layers: an input layer, a competitive layer, and a winner reverberation layer.

  In the input layer, units for inputting each component of the input vector constituting the time series data are arranged. In the present embodiment, a speed vector series generated from a series of coordinates indicating handwriting of handwritten numerals is used as time series data to be recognized. FIG. 2 shows a procedure for generating time-series data to be recognized in the present embodiment. As shown in the figure, when a user first writes a number on a tablet or the like, the coordinates of the writing position for each predetermined time are acquired. Then, by calculating the difference between the coordinates of the writing position that moves back and forth in time, a series of velocity vectors at the writing position can be obtained for each handwritten numeral. In the figure, a circle drawn on each handwritten numeral indicates a writing position for each predetermined time acquired by a tablet or the like. The graph shown on the right side of each handwritten numeral shows the transition of the value of each component in the vertical direction and the horizontal direction of the velocity vector at the writing position. As shown in the following equation (1), each handwritten numeral is represented by a series (time series data) of finite (eg, T) velocity vectors (input vectors), and each velocity vector is expressed by the following equation (2). As shown, it has a vertical component x1 (t) and a horizontal component x2 (t). Here, two units are arranged in the input layer, x1 (t) is input to one and x2 (t) is input to the other.

  N units (here, N = 6 × 6) are squarely arranged in the competitive layer, and each unit has a combined weight vector (first weight vector) Wj and an empirical weight vector (second weight vector). Uj is associated. When each input vector constituting the time series data is input to the input layer, one winner unit is selected from the units arranged in the competitive layer based on the coupling weight vector Wj and the experience weight vector Uj of each unit. Is done. The joint weight vector Wj is a vector having the same dimension as the input vector as shown in the following equation (3), and the empirical weight vector Uj is the same as the units arranged in the competitive layer as shown in the following equation (4). A vector having dimensionality.

  The same number of N units as the units arranged in the competitive layer are also arranged squarely in the winner reverberation layer. Each unit Aj of the winner reverberation layer corresponds to the unit Bj of the competitive layer. Each unit Aj of the winner reverberation layer is associated with a winner reverberation value Yj (t). The winner reverberation value Yj (t) increases when the competitive layer unit Bj is selected as the winner unit, and decreases with time. Accordingly, the winner reverberation vector Y (t) having the components of the winner reverberation values Y1 (t) to YN (t) shown in the following equation (5) indicates the history of the winner unit in the units arranged in the competitive layer.

  FIG. 3 is a diagram for explaining the update of the winner reverberation vector. In the figure, the bar drawn on the unit of the winner reverberation layer indicates the size of the winner reverberation value associated with the unit. As shown in the figure, when a unit having a competitive layer is selected as a winner unit at time 1, the winner reverberation value associated with that unit rapidly increases at time 2. Further, when another unit in the competitive layer is selected as a winner unit at time 2, the winner reverberation value associated with that unit rapidly increases at time 3. At this time, the winner reverberation value that rapidly increased at time 2 decreases at time 3 when time passes. Similarly, when another unit in the competitive layer is selected as a winner unit at time 3, the winner reverberation value associated with that unit rapidly increases at time 4. At this time, the winner reverberation value that rapidly increased at time 3 decreases at time 4 as time elapses. In addition, the winner reverberation value that rapidly increased at time 2 further decreases with the passage of time. Thus, the winner reverberation vector Y (t) includes information on the winner units before time t−1 and the order in which they were selected as the winner units.

  Specifically, the winner reverberation vector is updated according to the following equation (6). From this equation, the value range of Yj becomes the following equation (7). In addition, Zj (t) included in the first term on the right side of Expression (6) is a value of 1 when a competitive unit Bj is selected as a winner unit at time t as shown in Expression (8) below. Is a function that takes a value of 0 if not selected. Further, γ included in the right side of Expression (6) is a forgetting function that takes a range of the following Expression (9), and when a unit is selected as a winner unit, the corresponding winner reverberation value rapidly increases. And a parameter indicating the extent to which the winner finish value decreases with time.

  Here, a preparation process required for recognizing time-series data will be described. FIG. 4 is a flowchart showing the preparation process. As advance preparation, first, learning of the connection weight vector Wj associated with each unit Bj of the competitive layer of the self-organizing map is performed using the time series data for learning (S101). Further, the empirical weight vector Uj associated with each unit Bj is also learned using the time series data for learning (S102). Thereafter, the known time-series data is input to the learned self-organizing map, and the identification number of the winner unit corresponding to the last one of the input vectors constituting the time-series data is output (S103). Then, the value is recorded as determination data in association with the contents of the handwritten numerals represented by the time-series data input to the self-organizing map in S103 (S104).

  Next, processing for determining the contents of time-series data will be described. FIG. 5 is a flowchart showing the content determination processing of time series data. In this process, first, the time series data for recognition is input to the learned self-organizing map, and the identification number of the winner unit corresponding to the last one of the input vectors constituting the time series data is output (S501). ). Then, the content of the handwritten number represented by the time-series data for recognition is determined by comparing the identification number with the determination data, and the determination result is output (S502).

  FIG. 6 is a flowchart showing the learning process of the connection weight vector. The process shown in the figure shows the process of S101 in FIG. 4 in detail. In this process, first, the connection weight vector Wj (j = 1 to N) is initialized with a random number, for example (S201). Next, one randomly selected from the input vectors x (1) to x (T) constituting each of a large number of learning time-series data prepared in advance (here, x (t )) Is acquired (S202). Then, the similarity between the input vector x (t) and the coupling weight vector Wj (j = 1 to N) is calculated by the following equation (10), and the winner unit c is calculated by the following equation (11) using the value. elect. That is, the winner unit c has the highest similarity between the corresponding combination weight vector Wj and the input vector x (t) among the units Bj (j = 1 to N) arranged in the competitive layer. Further, a unit arranged in the vicinity of the winner unit c (within a predetermined vicinity range) in the competitive layer is also selected (S203).

  Then, the connection weight vector Wj associated with the winner unit c and its neighboring units is updated according to the following equation (12). Here, Wjnew is a weight vector Wj after update, and Wjold is a weight vector Wj before update. Α is a learning coefficient.

  Thereafter, it is determined whether there are any input vectors prepared in advance that have not yet been selected in S202 (S205). If there is such a thing, the processes of S202 to S204 are executed for those input vectors. Further, the processes of S202 to S205 are repeated a predetermined number of times (S206). At this time, the learning coefficient α is gradually reduced, and the predetermined neighborhood range is gradually narrowed. With the above learning process, the connection weight vector Wj associated with each unit Bj in the competitive layer corresponds to a velocity vector that frequently appears in handwritten characters.

  Next, FIG. 7 is a flowchart showing learning processing of the experience weight vector Uj. The process shown in the figure shows the process of S102 of FIG. 4 in detail. In this process, first, the experience weight vector Uj (j = 1 to N) is initialized, for example, with a random number (S301). Next, a time variable t is set to 1 (S302), and an input vector x (t) constituting time series data for learning is acquired (S303). Then, the temporary winner unit k is determined by the following equation (13) (S304). Here, in the following expression (13), the expression in parentheses on the right side indicates the similarity between the input vector x (t) and the coupling weight vector Wj. Further, a plurality of winner candidate units h that satisfy the following expression (14) are determined (S305). Where Φ is a given constant.

  Thereafter, the winner unit s is selected from the winner candidate units h by the following equation (15) (S306). That is, the similarity between the winner reverberation vector Y (t) and the experience weight vector Uj (j∈h) is calculated for all winner candidate units h by the following equation (16), and the similarity is maximized (that is, , The value of Equation (16) is minimized) is selected as the winner unit s. Then, the experience weight vector Us of the winner unit s is updated by the following equation (17). Here, Usnew is the updated experience weight vector Us, and Usold is the updated experience weight vector Us.

  Thereafter, the winner reverberation vector Y (t) is updated by the above equations (6) and (8) to calculate the winner reverberation vector Y (t + 1) at time t + 1 (S308). Then, t is incremented by 1 (S310) and S303 to S308 are repeated until the value of t is equal to or greater than T (the number of input vectors constituting learning time-series data). In this way, the empirical weight vectors Uj (j = 1 to N) are learned using the input vectors x (1) to x (T) constituting the time series data for learning in order. The above processing is similarly performed for time-series data corresponding to other handwritten numerals (S311), and both are repeated a predetermined number of times (S312). Through the above learning process, the experience weight vector Uj associated with each unit Bj of the competitive layer corresponds to the winner reverberation vector Y (t) when the unit Bj is selected as the winner unit.

  Here, a process for recognizing time-series data using the self-organizing map after learning the connection weight vector Wj and the experience weight vector Uj as described above will be described. FIG. 8 is a flowchart showing time-series data recognition processing, and shows details of the processing in S103 in FIG. 4 and S501 in FIG. In this process, first, a time variable t is set to 1 (S401), and an input vector x (t) constituting time series data for recognition is acquired (S402). Then, the temporary winner unit k is determined by the above equation (13) (S403). Further, a plurality of winner candidate units h that satisfy the above equation (14) are determined (S404).

  Thereafter, the winner unit s is selected from the winner candidate units h by the above equation (15) (S405). That is, the similarity between the winner reverberation vector Y (t) and the experience weight vector Uj (j∈h) is calculated by the above equation (16) for all winner candidate units h, and the similarity is maximized (that is, , The value of Equation (16) is minimized) is selected as the winner unit s.

  Next, the winner reverberation vector Y (t) is updated by the above equations (6) and (8) to calculate the winner reverberation vector Y (t + 1) at time t + 1 (S406). Then, t is incremented by 1 until the value of t becomes equal to or greater than T (the number of input vectors constituting the recognition time-series data) (S408), and S402 to S406 are repeated. In this way, the input vectors x (1) to x (T) constituting the time series data for recognition are sequentially input to the self-organizing map, and the winner unit s is sequentially determined. Then, the identification number s of the winner unit s obtained for the final input vector, that is, the input vector x (T) is output (S409).

  In the self-organizing map shown in FIG. 1, the connection weight vector Wj associated with each unit Bj in the competitive layer corresponds to a velocity vector that frequently appears in handwritten characters by prior learning. Further, the experience weight vector Uj associated with each unit Bj in the competitive layer corresponds to the winner reverberation vector Y (t) when the unit Bj is selected as a winner unit by prior learning. Therefore, the input vector x (t) constituting the time series data for recognition is input to the self-organizing map in time order, the similarity between the input vector x (t) and the combined weight vector Wj is high, and the winner finish The winning unit s corresponding to the final input vector x (T) can be recognized by sequentially determining the unit Bj of the competitive layer having a high similarity between the vector Y (t) and the experience weight vector Uj as the winning unit s. This corresponds to the contents of the handwritten numerals indicated by the time-series data input to. In this embodiment, the contents of the handwritten numerals indicated by the time-series data input for recognition are determined by checking the identification number s of the winner unit s against the determination data. As described above, according to this embodiment, the transition of the time-series data for recognition can be reflected in the data recognition, so that the recognition accuracy can be improved.

  FIG. 9 is a diagram illustrating a configuration of a time-series data recognition apparatus to which the above-described time-series data recognition method is applied. As shown in the figure, the time-series data recognition apparatus 10 includes an input vector receiving unit 12, a winner candidate determining unit 14, a combined weight vector storage unit 16, an experience weight vector storage unit 18, a winner determining unit 20, a winner reverberation vector. A storage unit 22, a winner reverberation vector update unit 24, a determination unit 26, and a determination data storage unit 28 are included. These elements are realized by causing various computers to execute the recognition program according to the present embodiment. The program may be stored in a computer-readable information storage medium such as a CD-ROM or DVD-ROM.

  First, the connection weight vector storage unit 16 stores the connection weight vector Wj in association with each unit Bj in the competitive layer, and the experience weight vector storage unit 18 stores the experience weight vector Uj in association with each unit Bj in the competition layer. To do. The values of the connection weight vector Wj and the experience weight vector Uj are obtained in advance by the processes of FIGS. The winner reverberation vector storage unit 22 stores the winner reverberation vector Y (t) at the current time t.

  The input vector receiving unit 12 receives input vectors x (1) to x (T) constituting time series data for recognition in time order. Then, the winner candidate determination unit 14 and the winner determination unit 20 sequentially input the similarity between the input vector x (t) and the combined weight vector Wj, and the similarity between the experience weight vector Uj and the winner reverberation vector Y (t). Based on the degree, one winner unit is selected from the units in the competitive layer. That is, based on the similarity between the sequentially accepted input vector x (t) and the coupling weight vector Wj associated with each unit Bj in the competitive layer, a part is selected from the units in the competitive layer. Based on the similarity between the experience weight vector Uj associated with the selected unit and the winner reverberation vector Y (t), one winner unit is selected from the selected units.

  Specifically, the winner candidate determination unit 14 substitutes the input vector x (t) received by the input vector reception unit 12 and the connection weight vector Wj read from the connection weight vector storage unit 16 into the above equation (13), Thereby, the temporary winner unit k is determined. Further, a plurality of winner candidate units h that satisfy the above equation (14) are determined. Then, the winner determination unit 20 reads the experience weight vector Uj related to the winner candidate unit h from the experience weight vector storage unit 18, and also reads the winner reverberation vector Y (t) from the winner reverberation vector storage unit 22, and sets them as the above formula ( Substituting into 15), the winner unit s is selected from the winner candidate units h.

  The winner reverberation vector update unit 24 updates the winner reverberation vector Y (t) according to the winner unit s selected by the winner determination unit 20. That is, the winner reverberation vector Y (t) is updated by the above equations (6) and (8), and the winner reverberation vector Y (t + 1) at time t + 1 is calculated.

  The determination data storage unit 28 inputs time-series data of known contents to the self-organizing map, and the identification number s of the winner unit s selected for the last input vector, and contents indicating the known contents Data (for example, numeric data indicating the contents of handwritten numerals) is stored in association with each other. The determination unit 26 obtains the identification number s of the winner vector s selected for the last one (x (T)) of the input vectors constituting the recognition time-series data, and uses the identification number s as the identification number s. The content data stored in the determination data storage unit 28 in association with each other is read and output as a determination result.

  According to the time-series data recognition device 10, it is possible to improve the recognition accuracy by reflecting the transition of the time-series data for recognition in the data recognition by relatively light information processing. Therefore, even when applied to a small computer such as a mobile phone or a personal digital assistant, handwritten characters can be recognized at high speed and with high accuracy.

It is a network block diagram of the self-organization map used in the recognition processing of time series data concerning this embodiment. It is a figure which shows the production | generation procedure of the time series data input into a self-organizing map for learning or recognition. It is a figure which shows the mode of the update of a winner reverberation vector. It is a flowchart which shows the process of prior preparation for recognition of time series data. It is a flowchart which shows the content determination processing of time series data. It is a flowchart which shows the learning process of a joint weight vector. It is a flowchart which shows the learning process of an experience weight vector. It is a flowchart which shows the recognition processing of time series data. It is a lineblock diagram of the time series data recognition device concerning this embodiment.

Explanation of symbols

  10 time-series data recognition device, 12 input vector reception unit, 14 winner candidate determination unit, 16 joint weight vector storage unit, 18 experience weight vector storage unit, 20 winner determination unit, 22 winner reverberation vector storage unit, 24 winner reverberation vector update Part, 26 determination part, 28 determination data storage part.

Claims (5)

A learning step that includes a competitive layer in which a plurality of units are arranged, and performs learning of a self-organizing map in which the first and second weight vectors are associated with each unit using time-series data for learning;
Recognizing time-series data for recognition using the self-organizing map,
The learning step includes
Sequentially receiving learning input vectors corresponding to the learning time-series data;
Selecting a part from the plurality of units based on the similarity between the learning input vector received sequentially and the first weight vector associated with each unit;
Bringing the first weight vector associated with the selected unit closer to the accepted learning input vector;
Accepting learning input vectors corresponding to the learning time-series data in chronological order;
Based on the similarity between the sequentially accepted learning input vector and the first weight vector associated with each unit, a part is selected from the plurality of units and the selected unit is selected. Based on the similarity between the second weight vector associated with the unit and the winner reverberation vector indicating the history of the winner unit with respect to the previously accepted learning input vector, Selecting one winner unit from the inside;
Bringing the second weight vector associated with the elected winner unit closer to the winner reverberation vector;
Updating the winner reverberation vector according to the selected winner unit;
The recognition step includes
Receiving recognition input vectors corresponding to the time-series data for recognition in chronological order;
Based on the similarity between the sequentially recognized recognition input vector and the first weight vector associated with each unit, a part is selected from the plurality of units and the selected unit is selected. The part selected based on the similarity between the second weight vector associated with a part of the unit and the winner reverberation vector indicating the history of the winner unit with respect to the previously recognized recognition input vector. Selecting one winner unit from among the units;
Updating the winner reverberation vector according to the selected winner unit;
Recognizing the time-series data for recognition by some or all of the selected winner units.
A time-series data recognition method characterized by the above. The time-series data recognition method according to claim 1,
The winner reverberation vector includes a winner reverberation value associated with each unit and updated as the unit is elected as a winner unit.
A time-series data recognition method characterized by the above. In the time series data recognition method according to claim 2,
The winner reverberation value decreases as the learning input vector or the recognition input vector is accepted,
A time-series data recognition method characterized by the above. A time-series data recognition device for recognizing time-series data using a self-organizing map having a competitive layer in which a plurality of units are arranged,
Weight vector storage means for storing the first and second weight vectors in association with each unit;
A recognition input vector receiving means for receiving a recognition input vector corresponding to time series data for recognition in time order;
The similarity between the recognition input vector and the first weight vector received sequentially, and the history of the winner unit for the recognition input vector previously received by the recognition input vector reception means and the second weight vector. Winner unit selection means for selecting one winner unit from the plurality of units based on the similarity to the winner reverberation vector shown;
Winner reverberation vector update means for updating the winner reverberation vector according to the selected winner unit;
Recognizing means for recognizing the time-series data for recognition by some or all of the selected winner units;
A time-series data recognition apparatus comprising: A program for recognizing time-series data using a self-organizing map having a competitive layer in which a plurality of units are arranged,
Weight vector storage means for storing the first and second weight vectors in association with each unit;
A recognition input vector receiving means for receiving recognition input vectors corresponding to time series data for recognition in time order;
The similarity between the recognition input vector and the first weight vector received sequentially, and the history of the winner unit for the recognition input vector previously received by the recognition input vector reception means and the second weight vector. Winner unit selection means for selecting one winner unit from the plurality of units based on the similarity to the winner reverberation vector shown;
The computer functions as a winner reverberation vector updating means for updating the winner reverberation vector according to the selected winner unit, and a recognition means for recognizing the time-series data for recognition by a part or all of the selected winner units. Program to let you.

Images