JP2008299640A - Pattern recognition apparatus, pattern recognition method, and program - Google Patents

Abstract

To provide a pattern recognition device, a pattern recognition method, and a program capable of automatically determining the number of states and the output distribution of each state to perform robust modeling of time series data.
A pattern recognition apparatus according to the present invention sequentially inputs a part or all of an input pattern as an input vector into a neural network, and is described by the input vector and a multidimensional vector arranged in the neural network. A template model generation unit 10 that generates a template model according to the feature quantity of the input pattern using the self-propagating neural network 12 that automatically increases the number of nodes based on the node, and the generated template model And a recognition unit 20 that recognizes the input pattern by matching the input pattern.
[Selection] Figure 1

Description

  The present invention relates to a pattern recognition apparatus, a pattern recognition method, and a program for performing pattern recognition using a self-propagating neural network.

  Time series pattern recognition and modeling are important fundamental technologies in various fields such as moving image processing, speech information processing, and DNA analysis. In general, the time-series pattern includes fluctuation in the feature space and expansion and contraction in the time direction. In order to recognize a time-series pattern robustly, it is necessary to construct a model and a learning device that can absorb these features. For this reason, model-based techniques are often used in which learning and recognition of time-series patterns is performed by having a model that holds a graph structure in advance.

  As a model-based technique, HMM (Hidden Markov Model) has been very successful as a standard technique in the field of speech recognition (see Non-Patent Document 1). In addition to speech recognition, HMM is used for speaker adaptation technology, speech synthesis technology, and the like, and has become a standard method in general speech information processing. HMM has been widely used for recognition of moving images and motions due to the success of this voice information processing and the support of statistical theory. As a method in speech recognition and moving image recognition, there are a method using a discrete HMM (Discrete HMM) and a method using a continuous distribution HMM (Continuous HMM) (see Non-Patent Documents 2 and 3). In addition, in order to accurately model the duration in each state, a segment model that explicitly has a duration distribution of each state has been proposed.

  The DP matching method, which is a kind of dynamic programming for HMMs, can calculate the distance between excessive time series patterns based on the local distance between feature parameters (each frame) in a short time. It is. DP matching is used for searching for time-series patterns in addition to voice recognition and motion recognition.

As a technique based on DP matching and HMM, a technique disclosed in Non-Patent Document 4 (hereinafter referred to as a stochastic DP method) has been proposed. In the stochastic DP method, a measure of probability is used as a measure of local distance in DP matching, and a path transition probability is used instead of a path cost. The stochastic DP method corresponds to one frame of a template pattern corresponding to one state, and corresponds to a left-to-right model having an HMM continuous output distribution with an increased number of states.
LR Rabiner, "A tutorial on hidden markov models and selected applications in speech recognition", Proc. IEEE, pp. 257-286 (1989). A. Wilson and A. Bobick, "Learning visual behavior for gesture analysis", Proc. IEEE International Symposium on Computer Vision, Vol. 5A Motion II (1995). R. Hamdan, F. Heits and L. Thoraval, "Gesture localization and recognition using probabilistic visual learning", Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Vol. 98-103 (1999). Seiichi Nakagawa, "Recognition of English consonants of unspecified speakers by stochastic DP method and statistical method", IEICE theory (D), vol.J70-D, no.1 (1987).

In the HMM, a Markov model having 3 to 5 states is often used for one phoneme for speech data because of the ease of parameter estimation. However, in such a small number of situations, an excessive time series pattern may not be accurately modeled.
Further, since the DP matching method uses the standard pattern itself as a model, it is difficult to model the feature space distribution in detail compared to the HMM.
On the other hand, the stochastic DP method is a method that utilizes both the advantage of DP matching and the robustness of the HMM, but a single multidimensional normal distribution is used for the output distribution of each state. In general, since the output distribution of each state differs depending on the number of dimensions and characteristics of the feature quantity, the problem is that the output distribution cannot be accurately approximated when such a single multidimensional normal distribution is used. There is.

  Thus, in the conventional pattern recognition model, it is necessary to determine the appropriate number of states and the output distribution of each state in advance, and the output distribution of each state is output using a single multidimensional normal distribution. There is a problem that the distribution cannot be approximated sufficiently.

  The present invention has been made in order to solve such problems, and is a pattern recognition apparatus and pattern recognition capable of automatically determining the number of states and the output distribution of each state and robustly modeling time-series data. It is an object to provide a method and a program.

  The pattern recognition apparatus according to the present invention sequentially inputs a part or all of an input pattern as an input vector to a neural network, and based on the input vector and a node described by a multidimensional vector arranged in the neural network. A template model generation unit that generates a template model according to the feature amount of the input pattern using a self-propagating neural network that automatically increases the node, and matches the generated template model with the input pattern And a recognition unit for recognizing the input pattern.

  As a result, the number of states in the template model and the output distribution of each state can be determined automatically, and the output distribution of each state can be approximated in detail, enabling robust modeling of time-series data. can do.

  The recognizing unit includes a likelihood calculating unit that calculates an output distribution of the state in the template model based on the node, and uses the calculated output distribution of the state in the template model. And the degree of coincidence with the input pattern may be calculated. Thereby, since the output distribution of each state in the template model can be approximated in detail, the time series data can be recognized with higher accuracy.

  Furthermore, the likelihood calculating unit includes a global likelihood calculating unit that calculates a global likelihood based on all nodes arranged in the self-propagating neural network, and the global likelihood is calculated from the global likelihood. An output distribution of states in the template model may be calculated. Thereby, by reflecting the feature quantity of the input pattern by the global likelihood, the output distribution of each state in the template model can be approximated in detail, and thus the time-series data can be recognized with higher accuracy.

  In addition, the likelihood calculating unit includes a local likelihood calculating unit that calculates a local likelihood based on a node belonging to the cluster for a cluster including nodes connected by edges, and the local likelihood The output distribution of the state in the template model may be calculated from the degree. Thus, by reflecting the feature quantity of the input pattern with local likelihood, the output distribution of each state in the template model even if the output distribution of each state cannot be approximated by a single multidimensional normal distribution Therefore, time series data can be recognized with higher accuracy.

  Furthermore, the likelihood calculation unit includes a global likelihood calculation unit that calculates a global likelihood based on all nodes arranged in the self-propagating neural network, and nodes connected by edges. A local likelihood calculating unit that calculates a local likelihood based on a node belonging to the cluster, and the state of the template model is determined from the global likelihood and / or the local likelihood. The output distribution may be calculated. As a result, even if the output distribution of each state cannot be approximated by a single multidimensional normal distribution by reflecting the feature quantity of the input pattern using the global likelihood and the local likelihood, the template Since the output distribution of each state in the model can be approximated in more detail, time-series data can be recognized with higher accuracy.

  In addition, the template model generation unit includes a standard pattern selection unit that selects a standard pattern of a class to which the input pattern belongs by matching between the input patterns, and sets the size of the input pattern to the size of the standard pattern. In other words, a part or all of each input pattern corresponding to each frame of the standard pattern may correspond to each state in the template model. Accordingly, the number of states in the template model can be determined according to the number of frames of the standard pattern without previously determining the number of states in the template model.

  Furthermore, the template model generation unit is configured to determine the number of input patterns of the input pattern and the feature amount of the input pattern with respect to a set of elements including part or all of the input patterns corresponding to the states in the template model. The element set may be input to the self-propagating neural network based on the number of dimensions. As a result, even if the number of elements in the element set corresponding to one state is small, it is necessary to approximate the distribution by determining the number of elements according to the number of input patterns and the number of dimensions of the input patterns. A large number of elements can be secured and input, and a reduction in recognition accuracy can be prevented.

  In addition, when the template model generation unit sequentially adds and inputs input patterns, the template model generation unit outputs the state of the template model corresponding to the class to which the sequentially added input pattern belongs. The distribution may be updated according to the sequentially added input pattern. As a result, additional input patterns can be easily learned, and a large amount of data is not required in advance, and recognition accuracy is improved based on a small amount of data given sequentially. be able to.

  Furthermore, the matching process may be performed using a DP matching method. Thereby, a matching process can be performed efficiently.

  The self-propagating neural network is configured such that when an edge is connected between a node having a weight vector closest to the input vector and a node having a second closest weight vector, the node of interest and other nodes A similarity threshold of the node of interest calculated based on the distance between, and an interclass node insertion unit that inserts the input vector as a node based on the distance between the input vector and the node of interest; and A weight vector updating unit that updates a weight vector corresponding to a node closest to the input vector and a weight vector corresponding to a node directly connected to the node by an edge so as to be closer to the input vector. May be. As a result, since the number of nodes to be inserted can be autonomously managed based on the similarity threshold, a new input pattern that is sequentially input can be created without determining the number of nodes in advance. You can learn more without breaking your knowledge.

  Furthermore, the inter-class node insertion unit sets a node having a weight vector closest to the input vector to be input as a first winner node and a node having a weight vector closest to the second as a second winner node. When an edge is connected between the 1 winner node and the 2nd winner node, if there is a node that is directly connected to the node of interest by the edge, the node is directly connected. The distance between the nodes having the maximum distance from the node of interest is the similarity threshold, and if there is no node directly connected to the node of interest by the side, the attention is paid A similarity threshold calculation unit that calculates a distance between nodes having a minimum distance from the node as the similarity threshold; the input vector; and the first winner no Whether the distance between the input vector and the second winner node is greater than the similarity threshold of the first winner node, and whether the distance between the input vector and the second winner node is greater than the similarity threshold of the second winner node And a node insertion unit that inserts the input vector as a node at the same position as the input vector based on the similarity threshold determination result. Thus, according to the similarity threshold that changes according to the input vector, the number of nodes to be inserted can be autonomously managed, so that the number of nodes is sequentially input without determining in advance. Additional input patterns can be learned without breaking existing knowledge.

  Further, the self-propagating neural network includes an edge deleting unit that deletes the edge based on the age of the edge associated with the edge, and an edge that is directly connected to the noticed node. And a node deletion unit that deletes the node of interest based on the number of nodes. Thereby, a node generated from an unnecessary input pattern can be effectively and dynamically deleted.

  Furthermore, the self-propagating neural network may have a single layer structure. As a result, non-patent literature: F. Shen and O. Hasegawa, "An Incremental Network for On-line Unsupervised Classification and Topology Learning," Neural Networks, vol. 19, pp. 90-106, 2006. Compared to Self-Organizing Incremental Neural Network (hereinafter referred to as “SOINN”), additional learning can be executed without specifying the timing for starting the second layer learning.

  The pattern recognition method according to the present invention sequentially inputs a part or all of an input pattern as an input vector to a neural network, and based on the input vector and a node described by a multidimensional vector arranged in the neural network. A template model generating step for generating a template model according to the feature quantity of the input pattern using a self-propagating neural network that automatically increases the node, and matching the generated template model with the input pattern And a recognition step for recognizing the input pattern.

  The program according to the present invention causes a computer to execute information processing as described above.

  According to the present invention, it is possible to provide a pattern recognition device, a pattern recognition method, and a program that can automatically determine the number of states and the output distribution of each state and perform robust modeling of time-series data. it can.

  Hereinafter, specific embodiments to which the present invention is applied will be described in detail with reference to the drawings. In addition, in each drawing, the same code | symbol is attached | subjected to the same element, and duplication description is abbreviate | omitted as needed for clarification of description.

Embodiment 1 of the Invention
In the first embodiment, the present invention is applied to a pattern recognition apparatus that recognizes time-series patterns. The pattern recognition device uses a self-propagating neural network, which is an online unsupervised learning method, to generate a template model according to the feature quantity of the time series pattern, and then matches the generated template model with the time series pattern. Recognize sequence patterns.
By using SOIN, which will be described later, as the self-propagating neural network, the output distribution of each state in the template model can be automatically determined and approximated in detail. Here, the pattern recognition apparatus inputs a part or all of the time-series pattern to the SOINN, calculates the likelihood based on the phase structure (a set of nodes and sides) formed in the SOINN, and outputs the state Approximate the distribution. By calculating the global likelihood and the local likelihood as the likelihood, it is possible to express the output distribution of the state according to the feature quantity of the time series pattern in detail, so that the time series pattern can be robustly modeled .

  FIG. 1 is a block diagram showing a pattern recognition apparatus 1 according to Embodiment 1 of the present invention. The pattern recognition apparatus 1 includes a template model generation unit 10 and a recognition unit 20, and includes a training pattern database (DB) 31 that stores training patterns for learning and a template model that stores template models generated as a result of learning. A learning result database (DB) 32 is connected.

  The template model generation unit 10 includes a standard pattern selection unit 11 and a SOINN 12. The SOIN 12 as a self-propagating neural network includes an interclass node insertion unit 121, an edge deletion unit 122, a node deletion unit 123, and a weight vector update unit 124. The recognition unit 20 includes a likelihood calculation unit 21. The likelihood calculating unit 21 includes a global likelihood calculating unit 211 and a local likelihood calculating unit 212.

Next, each block will be described in detail below.
The template model generation unit 10 generates a model of each class (hereinafter referred to as a template model) according to the feature amount of the input pattern using the SOIN 12 described later. More specifically, first, the standard pattern selection unit 11 to be described later selects a central standard pattern from the training data, and performs DP matching between this standard pattern and other training patterns. Normalize the training pattern to the time series length of the standard pattern. Next, a set of data consisting of a part or all of the input pattern corresponding to each frame of the standard pattern is made to correspond to one state in the template model, and the distribution of this set of data is approximated by SOIN 12. As a result, the template model holds the number of states as many as the number of frames of the standard pattern, and the output distribution of each state can be approximated by the SOINN 12. Hereinafter, this method is referred to as the SOIN-DP method.

The standard pattern selection unit 11 selects a standard pattern of a class to which the input pattern belongs by matching between the input patterns. Here, the DP matching method is used for the matching process. By using the DP matching method, the matching process can be performed efficiently. By performing the DP matching, the accumulated distance D between the two patterns X and Y (X, Y), and optimal correspondence _j = w i between patterns (i = 1,2, ···, I ) the Obtainable.

Hereinafter, the DP matching method will be briefly described.
In the first embodiment, the time series pattern _X = the number of frames _{I {x 1, x 2,} ···, x i, ···, x I}, and when the frame number J-series pattern Y = { Considering DP matching with y ₁ , y ₂ ,..., y _j ,..., y _J }, the cumulative distance D (X, Y) of these two time series patterns is calculated. Here, i and j indicate the frame numbers of the time series patterns X and Y, respectively. In addition, the feature vector x _i of each frame of X is referred to as an element of the i-th frame of X or an i-th element. In the first embodiment, the cumulative distance D (X, Y) between the time series pattern X and the time series pattern Y is calculated using the following symmetric recurrence formula.
Then, the cumulative distance D (X, Y) is calculated based on the following formula using the recurrence formula shown in Equation 2 above.

Thus, according to DP matching, the cumulative distance D (X, Y) can be calculated by incrementally repeating the operation of accumulating the current local distance in the cumulative distance. Further, the DP matching, the i-th X element _{x i} and Y in (frame) the j-th element _{y j} and the optimum correspondence _j = w i (i = 1, 2, · · ·, I) can be obtained.
As a recurrence formula used for DP matching, in addition to the symmetric recurrence formula, there is an asymmetric recurrence formula shown below.

  The SOINN 12 is a self-propagating neural network that automatically increases the number of nodes based on the input vector and the nodes arranged in the neural network. In the first embodiment, the first layer of the SOINN described below is used. Use to learn. By adopting a one-layer structure, it is possible to execute additional learning without specifying the timing for starting learning of the second layer, as compared to SOIN. Also, the present invention is not limited to SOIN, and can be a single-layer structure such as Enhanced-SOINN (hereinafter referred to as E-SOINN), which will be described later. Can be executed.

In the following, first, SOIN, which is a conventional technique, will be briefly described, and then SOIN 12 according to the first embodiment will be described. SOINN is a technique disclosed in Non-Patent Literature: Fritzke.B, “A growing neural gas network learns topologies,” Advances in neural information processing systems (NIPS), pp 625-632, 1995. This is an unsupervised additional learning technique called a so-called self-propagating neural network, which is an extension of GNG). A network that expresses the distribution of input data can be additionally constructed by learning input vectors sequentially while self-propagating nodes. SOIN has the following four advantages.
(1) A new cluster can be constructed by additionally learning an input vector of a newly input unknown class without destroying a previously learned cluster.
(2) Noise independent of input data can be effectively and dynamically removed.
(3) For unsupervised data given sequentially, a network that expresses the phase structure can be autonomously constructed.
(4) The input vector can be approximated without determining the number of nodes in advance.

  FIG. 2 is a flowchart for explaining a learning process by SOIN which is a conventional technique. Hereinafter, the SOIN process will be briefly described with reference to FIG. Here, SOINN has a two-layer network structure and performs the same learning process in the first layer and the second layer. In addition, the SOIN uses the learning result that is the output of the first layer as an input vector to the second layer.

S101: An input vector is given to SOINN.
S102: A node closest to the given input vector (hereinafter referred to as a first winner node) and a second closest node (hereinafter referred to as a second winner node) are searched.
S103: Based on the similarity threshold of the first winner node and the second winner node, it is determined whether or not the input vector belongs to the same cluster as at least one of these winner nodes. Here, the node similarity threshold is calculated based on the idea of the Voronoi region. In the learning process, the position of the node gradually changes to approximate the distribution of the input vector, and the Voronoi region also changes accordingly. That is, the similarity threshold value adaptively changes according to the change in the position of the node.
S104: If the result of determination in S103 is that the input vector belongs to a different cluster from the winner node, the node is inserted at the same position as the input vector, and the process proceeds to S101 to process the next input vector. This insertion is called interclass insertion.

S105: On the other hand, when the input vector belongs to the same cluster as the winner node, an edge is generated between the first winner node and the second winner node, and the nodes are directly connected by the edge.
S106: Update the weight vectors of the first winner node and the nodes directly connected to the first winner node by edges.
S107: The edge generated in S105 has an age, and an edge having an age exceeding a preset threshold is deleted. In online learning in which input vectors are given sequentially, the position of the node always changes gradually, so that the adjacency constructed by the initial learning may not be established by subsequent learning. For this reason, by configuring the sides that are not updated even after a certain period of time so that the age of the sides becomes high, it is possible to delete the sides that are not necessary for learning.

  S108: It is determined whether the total number of inputs of the input vector is a preset multiple of λ. As a result of the determination, if the total number of input vectors is not a multiple of λ, the process returns to S101 to process the next input vector. On the other hand, when the total number of input vectors is a multiple of λ, the following processing is executed.

  S109: A node having the maximum local accumulated error is searched, and a new node is inserted in the vicinity of the node (this insertion is called intra-class insertion). Then, based on the error radius indicating the average error of the node, it is determined whether or not the node insertion is successful. Here, the local accumulated error is calculated by accumulating the error of the node according to the input of the input vector as the error of the node, which is the distance difference between the node and the input vector. The error radius is calculated based on the error of the node and the number of times the node has become the first winner.

S110: When it is determined that the node insertion by the intra-class insertion is successful, the node inserted by the intra-class insertion and the node having the maximum local accumulated error are directly connected by the edge. On the other hand, if it is determined that node insertion by intra-class insertion has failed, the node inserted by intra-class insertion is deleted, and the process proceeds to S111.
S111: The noise node is deleted based on the number of adjacent nodes and the number of times the node becomes the first winner.
Here, the adjacent node indicates a node that is directly connected to the node by a side, and a node whose number of adjacent nodes is 1 or less is a deletion target. Also, the cumulative number of times of becoming the first winner is compared with a threshold value calculated using a preset parameter c, and a node whose first winner cumulative number is less than the threshold value is targeted for deletion.

S112: It is determined whether or not the total number of inputs of the input vector is a preset multiple of LT. If it is determined that the total number of input vectors is not a multiple of LT, the process returns to S101 to process the next input vector. On the other hand, when the total number of input vectors is a multiple of LT, the following processing is executed.
S113: It is determined whether or not learning for the first layer is to be ended. As a result of the determination, when the process proceeds to the learning of the second layer, the process proceeds to S101 and the node that is the learning result of the first layer is input as an input vector to the second layer. However, when additional learning is performed, the previous learning result remaining in the second layer is deleted, and then the second layer learning is started.
When the number of inputs to the second layer is a multiple of the preset number of times LT and the learning of the second layer is terminated, the nodes are classified into different classes, and the number of classes and representative prototype vectors of each class are Output and stop. Here, the prototype vector corresponds to a node weight vector.

Here, an experiment performed using an artificial data set to verify the function of SOIN is shown.
FIG. 3 is an image showing two-dimensional artificial data input to the SOINN. The input data set shown in FIG. 3 is composed of a total of five signal sources (classes) of two Gaussian distributions, two concentric circles, and a sine curve. In addition, assuming an actual environment, 10% uniform noise was added to signals generated from five signal generation sources. The artificial data set shown in FIG. 3 was additionally input online to SOIN to perform unsupervised class classification.

  FIG. 4 is an image showing an output result when the two-dimensional artificial data shown in FIG. 3 is additionally inputted to SOINN. As shown in FIG. 4, the SOIN can remove noise included in the input data, and can correctly extract the number of classes of the input data and its topology (phase structure).

  In this way, SOINN can learn non-stationary inputs by autonomously managing the number of nodes, and extracts the appropriate number of classes and topological structure even for classes with complex shapes in the distribution. It has many advantages such as being able to. As an application example of SOIN, for example, in pattern recognition, after learning a hiragana character class, it is possible to additionally learn a katakana character class. Further, since the number of nodes can be automatically increased by using SOINN as a self-propagating neural network, it is not limited to a stationary environment in which input vectors are randomly given from the input vector space. It is also possible to cope with a non-stationary environment in which the class to which the input vector belongs is switched and the input vector is randomly given from the switched class.

Next, the SOINN 12 according to the first embodiment will be described.
The SOIN 12 includes an interclass node insertion unit 121, an edge deletion unit 122, a node deletion unit 123, and a weight vector update unit 124.
The interclass node insertion unit 121 includes a similarity threshold calculation unit, a similarity threshold determination unit, and a node insertion unit. The inter-class node insertion unit 121 sets a node having a weight vector closest to the input vector to be input as a first winner node and a node having a weight vector closest to the second as a second winner node. When an edge is connected between two winner nodes, a node is inserted as described below.

  First, when there is a node that is directly connected to the node of interest by a side with respect to the node of interest, the similarity threshold calculation unit calculates the distance from the node of interest among the directly connected nodes. If the distance between the nodes that is the maximum is the similarity threshold, and there is no node that is directly connected to the target node by the side, the distance between the nodes that has the minimum distance from the target node is the similarity. Calculate as the threshold.

Next, the similarity threshold determination unit determines whether the distance between the input vector and the first winner node is greater than the similarity threshold of the first winner node, and the distance between the input vector and the second winner node is the second winner. It is determined whether it is larger than the similarity threshold of the node.
Next, the node insertion unit inserts the input vector as a node at the same position as the input vector based on the similarity threshold determination result.

  In this way, according to the similarity threshold that changes according to the input vector, the number of inserted nodes can be autonomously managed, so that the number of nodes can be sequentially determined without being determined in advance. Additional input associations can be learned without breaking existing knowledge.

The weight vector update unit 124 updates the weight vector corresponding to the node closest to the input vector and the weight vector corresponding to the node directly connected by the node and the edge so as to be closer to the input vector.
The side deleting unit 122 deletes the side based on the age of the side associated with the side. The node deletion unit 123 deletes the node of interest based on the number of sides directly connected to the node of interest for the node of interest. As a result, an erroneously generated edge can be appropriately deleted. A node that does not have an edge indicates that the frequency of input close to the connection weight vector of the node is extremely low, and the information held by the node is considered to be noise unrelated to the data to be learned. It is because it can do.

FIG. 5 is a flowchart for explaining learning processing by the SOINN 12. Hereinafter, the processing of the SOINN 12 will be described with reference to FIG.
S201: The SOINN 12 acquires two input vectors, initializes the node set A as a set including only two nodes corresponding to them, and stores the result in the temporary storage unit. Also, the edge set C⊂A × A is initialized as an empty set, and the result is stored in the temporary storage unit.

S202: The SOINN 12 inputs a new input vector ξ and stores the result in the temporary storage unit.
S203: The SOINN 12 determines the first winner node a ₁ having the weight vector closest to the input vector ξ and the second winner node a ₂ having the second closest weight vector for the input vector and the node stored in the temporary storage unit. Search and store the result in the temporary storage.
S204: interclass node insertion unit 121, an input vector stored in the temporary storage unit, node, the node similarity threshold, the distance between the input vector ξ and the first winning node a ₁ is first winning node a ₁ It is determined whether or not it is greater than the similarity threshold T ₁ and whether the distance between the input vector ξ and the second winner node a ₂ is greater than the similarity threshold T ₂ of the second winner node a ₂ , and the result Is stored in the temporary storage unit. Here, the similarity threshold T ₂ of the similarity threshold T ₁ and the second winning node a ₂ of the first winning node a ₁ stored in the temporary storage unit is calculated in the same manner as SOINN, the result is temporarily stored Stored in the department.

S205: As a result of the determination in S204 stored in the temporary storage unit, the distance between the input vector ξ and the first winner node a ₁ is greater than the similarity threshold T ₁ of the first winner node a ₁ , or the input vector ξ when the distance between the second winning node a ₂ is greater than the similarity threshold T ₂ of the second winning node a ₂ are, inter-class node insertion unit 121, the input vector and node stored in the temporary storage unit, an input The vector ξ is inserted as a new node i at the same position as the input vector ξ, and the result is stored in the temporary storage unit.
S206: On the other hand, as a result of the determination in S204 stored in the temporary storage unit, the distance between the input vector ξ and the first winner node a ₁ is equal to or less than the similarity threshold T ₁ of the first winner node a ₁ and the input When the distance between the vector ξ and the second winner node a ₂ is equal to or less than the similarity threshold T ₂ of the second winner node a ₂ , the SOIN 12 stores the nodes stored in the temporary storage unit and the edges between the nodes. It is determined whether or not an edge is connected between the first winner node a ₁ and the second winner node a ₂ , and the result is stored in the temporary storage unit.

S207: As a result of the judgment in S206 stored in the temporary storage unit, when connecting to generate edges between the first winning node a ₁ and the second winning node a _2, SOINN12 is stored in the temporary storage unit For the nodes and the sides between the nodes, the sides are connected between the first winner node and the second winner node, and the result is stored in the temporary storage unit. Then, the SOIN 12 sets the side age of the side and the age of the side stored in the temporary storage unit to 0 for the newly generated side and for the side if the side has already been generated between the nodes. Set and store the result in the temporary storage unit, increment the age of the side directly connected to the first winner node a ₁ (increase by 1), and store the result in the temporary storage unit.
On the other hand, the result of determination in S206 stored in the temporary storage unit, when not connected to edges between the first winning node a ₁ and the second winning node a _2, although the process proceeds to S208, already between nodes If the edge has been generated, the SOIN 12 deletes the edge between the first winner node a ₁ and the second winner node a ₂ for the node stored in the temporary storage unit and the edge between the nodes, and as a result Is stored in the temporary storage unit.
Next, the SOIN 12 increments (accumulates by 1) the cumulative number of times M _a1 that the first winner node a ₁ stored in the temporary storage unit has become the first winner node, and stores the result in the temporary storage unit.

S208: weight vector updating unit 124, the weight vector of the stored node and the node in the temporary storage unit, respectively the input vector the weight vector of the weight vector of the first winning node a ₁ and first winning node a ₁ neighbor nodes Update it so that it is closer to ξ, and store the result in the temporary storage. Here, for calculation of the update amount of the weight vector, M _a1 stored in the temporary storage unit is used as t.
S209: side deleting unit 122 stores the stored edge in the temporary storage unit, delete the edges with a exceeds a preset threshold age _t stored in the temporary storage unit age, the result in the temporary storage unit To do. The age _t is used to delete a side that is erroneously generated due to the influence of noise or the like. By setting a small value for age _t , edges can be easily deleted and the effect of noise can be prevented. However, if the value is extremely small, edges are frequently deleted and the learning result becomes unstable. Become. On the other hand, if excessively set to a large value age _t, it can not be removed edges generated by influence of noise appropriately. Taking these into account, the parameter age _t is calculated in advance by experiments and stored in the temporary storage unit.

S210: SOIN 12 determines whether or not the total number of input vectors ξ stored in the temporary storage unit is a multiple of λ set in advance and stored in the temporary storage unit. The determination is made, and the result is stored in the temporary storage unit. As a result of the determination stored in the temporary storage unit, if the total number of input vectors is not a multiple of λ, the process returns to S202 to process the next input vector ξ. On the other hand, when the total number of input vectors ξ is a multiple of λ, the following processing is executed. Note that λ is a period for deleting a node regarded as noise. Although it is possible to frequently perform noise processing by setting a small value to λ, if the value is extremely small, nodes that are not actually noise are erroneously deleted. On the other hand, if an extremely large value is set to λ, a node generated due to noise cannot be removed appropriately. In consideration of these, the parameter λ is calculated in advance through experiments and stored in the temporary storage unit.
S211: The node deletion unit 123 deletes a node regarded as a noise node for the node stored in the temporary storage unit, and stores the result in the temporary storage unit.

S212: The SOINN 12 determines whether or not the total number of input vectors ξ stored in the temporary storage unit is a multiple of a preset LT and the result Store in the temporary storage. If the total number of input vectors is not a multiple of LT as a result of the determination stored in the temporary storage unit, the process returns to S202 to process the next input pattern ξ. On the other hand, when the total number of input vectors ξ is a multiple of LT, the following processing is executed.
S213: The SOINN 12 outputs the node stored in the temporary storage unit as a prototype. After completing the above processing, learning by the SOINN 12 is stopped.

The recognition unit 20 recognizes which class the input pattern belongs to by matching the input pattern with a template model composed of training patterns of each class. The DP matching between the template model TM _c and the input pattern IP class c using a symmetric recurrence formula, using the following recurrence formula. The likelihood C (x _i , S _j ) indicates the likelihood of the i-th element x _i of the input pattern IP with respect to the j-th state S _j of the template model TM _c. Calculated.
Here, the same symmetric recursion formula as in Equation 2 was used. Thereby, in the recognition experiment of real data, recognition accuracy can be improved compared with the case where an asymmetric type recurrence formula is used.

Then, the recognizing unit 20 performs DP matching so that the sum of the likelihoods C (x _i , S _j ) is maximized. That is, in Equation 2, min is used to minimize the accumulated distance g (i, j), but in Equation 5, max is used to maximize Q (i, j).
As a result of the DP matching, a cumulative matching degree E (IP, TM _c ) between the template model TM _c and the input pattern IP is calculated based on the following equation. I _IP represents the time series length of the input pattern IP, and J _c represents the time series length of TM _c .
Therefore, the recognition unit 20 recognizes the input pattern based on the following expression. Here, the following expression is a function that outputs the class number of the template model having the largest input pattern IP and cumulative match E (IP, TM _c ). In this case, the attribute class of the input pattern IP is c ^* . Recognize that there is.

The likelihood calculating unit 21 includes a global likelihood calculating unit 211 and a local likelihood calculating unit 212, and calculates an output distribution of states in the template model based on the nodes. This will be described in detail below.
The likelihood calculating unit 21 calculates the likelihood C (x _i , S _j ) based on the nodes arranged in the SOINN 12 and the cluster composed of nodes connected by the sides. Specifically, based on the following equation, global likelihood _{log (P whole (x i |} S j)) and local likelihood _P class _| using _(x i _{U jk),} the likelihood C ( x _i , S _j ) is calculated. FIG. 6 is a diagram showing a plurality of inner classes (here, Class 1 to 3) that are clustered by the SOINN 12 and exist in the SOINN 12. As shown in FIG. One cluster generated by the SOINN 12 is defined as an inner class. The inner class is represented by a node reference vector (prototype vector group).
Here, w _k is calculated based on the following equation. _{Incidentally, N all} denotes the total number of all the nodes in the SOINN12 state _{S j,} K is the number of internal classes in SOINN12 state _{S j.}

  As a result, even if the output distribution of each state cannot be approximated by a single multidimensional normal distribution by reflecting the feature quantity of the input pattern using the global likelihood and the local likelihood, the template Since the output distribution of each state in the model can be approximated in more detail, time-series data can be recognized with higher accuracy. In addition, by reflecting the global likelihood and the local likelihood according to the input pattern, it is possible to respond more flexibly to the output distribution of each state.

The global likelihood calculating unit 211 calculates the global likelihood based on all the nodes arranged in the SOINN 12. Specifically, all nodes existing in the SOINN 12 in the j-th state S _j are used to calculate the global likelihood. The global likelihood is calculated by approximating all nodes existing in the state SOINN 12 with a probability density function of a multidimensional normal distribution, and the occurrence probability _Pwhole (x _i | S _j ) from this density function. The occurrence probability P _whole (x _i | S _j ) is calculated based on the following equation. Here, μ _i is an average vector of all nodes existing in the SOIN 12 in the state S _j , and Σ _j is a covariance matrix. These two parameters are calculated by maximum likelihood estimation. The log likelihood log (P _whole (x _i | S _j )) of the occurrence probability P _whole (x _i | S _j ) is defined as the global likelihood.

The local likelihood calculating unit 212 calculates a local likelihood for a plurality of clusters existing in the SOINN 12 based on nodes belonging to the cluster. The local likelihood is calculated using information of a plurality of inner classes clustered by the SOINN 12. For example, each inner class (Class 1 to Class 3) shown in FIG. 6 is approximated by a kernel function based on a multivariate normal distribution. The reason for using the kernel function here is that each inner class has a small number of nodes (at least two), and it is difficult to estimate a multidimensional normal distribution from such a small number of data. is there. The kth inner class in the SOINN 12 is defined as U _jk, and the local likelihood P _class (x _i | U _jk ) estimated from U _jk is calculated based on the following equation.
Here, x _jk is an average vector of all nodes existing in the inner class U _jk in the SOINN 12, and h _jk is a parameter indicating the size of the region of the kernel function, and is calculated based on the following equation. Incidentally, _{a t} denotes the position vector of the node t of the internal class _{U jk, N jk} indicates the total number of nodes included within the class _{U jk.}

  The pattern recognition apparatus 1 as described above can be realized by a computer such as a dedicated computer or a personal computer (PC). However, the computer does not need to be physically single, and a plurality of computers may be used when performing distributed processing.

  FIG. 7 is a diagram illustrating an example of a system configuration for realizing the pattern recognition apparatus 1 according to the first embodiment. As shown in FIG. 7, the computer 40 includes a CPU 41 (Central Processing Unit), a ROM 42 (Read Only Memory), and a RAM 43 (Random Access Memory), which are connected to each other via a bus 44. Although explanation of OS software for operating the computer is omitted, it is assumed that a computer for constructing the pattern recognition apparatus 1 is also provided.

  An input / output interface 45 is also connected to the bus 44. The input / output interface 45 includes, for example, an input unit 46 including a keyboard, a mouse, and a sensor, a display including a CRT and an LCD, an output unit 47 including a headphone and a speaker, a storage unit 48 including a hard disk, A communication unit 49 including a modem and a terminal adapter is connected.

  The CPU 41 executes various processes according to various programs stored in the ROM 42 or various programs loaded from the storage unit 48 to the RAM 43. In the present embodiment, for example, a point play model generation process and a recognition process are executed. The RAM 43 also appropriately stores data necessary for the CPU 41 to execute various processes. The communication unit 49 performs, for example, communication processing via the Internet (not shown), transmits data provided from the CPU 41, and outputs data received from a communication partner to the CPU 41, the RAM 43, and the storage unit 48. The storage unit 48 communicates with the CPU 41 to save and erase information. The communication unit 49 also performs communication processing of analog signals or digital signals with other devices. The input / output interface 45 is also connected to a drive 50 as required. For example, a magnetic disk 501, an optical disk 502, a flexible disk 503, or a semiconductor memory 504 is appropriately mounted, and a computer program read from them is required. Is installed in the storage unit 48 accordingly.

  Next, template model generation processing and recognition processing according to the first embodiment will be described. FIG. 8 is a flowchart showing an outline of template model generation processing by the pattern recognition apparatus 1. Hereinafter, template model generation processing by the pattern recognition apparatus 1 will be described with reference to FIG. The template model generation unit 10 generates a template model belonging to class C as follows. Here, the template model generation unit 10 generates a template model for each class.

S301: N training patterns are given from the training pattern DB 31, and the standard pattern selection unit 11 selects one standard pattern from the training pattern group. More specifically, DP matching is performed between a certain training pattern P _m included in class C and another training pattern excluding P _m included in class C. This processing is performed by all combinations (brute force) of training patterns in class C. The sum of accumulated distances between patterns obtained from the DP matching result is calculated, and a pattern having the smallest accumulated distance is selected as a standard pattern. That is, a standard pattern is selected based on the following formula. Here, arg outputs the number m ^{* of the} training pattern that minimizes the sum of the cumulative distances between the training patterns.
In this way, the m ^* th training pattern of class C is selected as the standard pattern P ^{* that} is the center of the template model. Note that the number of frames T ^* of P ^* is the time series length of the template model.

S302: The training pattern frame data associated in the DP matching is input to the SOINN 12 as a state in the template model. More specifically, the input to the SOINN 12 can be realized using the result of DP matching. That is, as a result of DP matching between the standard pattern P ^* and the other N-1 training patterns, the time series length of all other training patterns is normalized to the time series length of the standard pattern P ^*. . Further, the correspondence between each element of the standard pattern P ^* and each element of the N−1 training patterns is obtained. Based on the result of such DP matching, an element (vector) group having a correspondence relationship is input to each SOIN 12 space (each state) as described below.

First, the j-th element of the standard pattern P ^* is p ^* _j , the i-th (i-th frame) element of the training pattern P _n (nεC) is p ⁿ _i, and p ^* _j and p ⁿ An optimal association w ⁿ with _i is defined as follows.
According to the above equation 14, the i-th element of the training pattern is distributed to the j-th state (SOINN 12 space). Such dispensing operation, and the standard pattern, after between other N-1 pieces of exercise pattern, the N-1 an optimal path ^{w n (n = 1, ···} , N-1) Can be obtained. According to the N-1 optimum paths, each element of the training pattern is distributed to each state.

Here, in the first embodiment, an element set group distributed in a certain time range (between states) is input to one SOINN 12. Specifically, an element set group distributed to the jth state is defined as Z ^* _j as shown in the following formula, and an element set from Zj to Zj _{+ L−1} is defined as the _jth state (SOINN12 ).
Note that L is a parameter, and this parameter is defined as the number of segments. The parameter setting method will be described later. Here, the number of states of the template model can be determined as T ^* −L−1 by using the Segment number L and the time series length T ^* of the standard pattern. Thus, the number of states in the template model can be determined according to the number of frames of the standard pattern without previously determining the number of states in the template model.

FIG. 9 is a diagram showing how a training pattern is input to the SOINN 12. In FIG. 9, Criterion Data indicates a standard pattern, and Data 1 to 3 indicate training patterns. Each block of Data and Criterion Data indicates an element vector at each time (frame). The same color part in each block indicates the corresponding part in the optimum route after DP matching. For example, the elements corresponding to the elements in the first frame of Criterion Data are the first and second elements of Data1, the first element of Data2, and the first element of Data3. These element sets are Z ₁ and Become. The solid line is a line connecting the corresponding element set Z _1. The broken lines indicate element set of _{Z 2,} the element set of _{Z 3,} respectively. Element set group ^Z _{* 1} from Z ₁ to _{Z L} are input to state 1 (SOINN12) in the template model class C.

S303: The SOINN 12 performs learning based on the frame data. In each state j, an element set group Z ^* _j as frame data is input to the SOINN 12 space. Here, since SOIN 12 is a technique that enables online learning, when the element set group Z ^* _j is input, each element of the element set group Z ^* _j is input at random one by one. When a randomly input vector is input to the SOINN 12 space, generation and deletion of nodes and edges are repeated in the SOIN 12 space, and finally a plurality of representative node sets (clusters) are formed. A learning result by the SOINN 12 is stored in the template model learning result DB 32. As described above, the likelihood calculating unit 21 estimates the output distribution of the state from the plurality of representative node sets formed in this way.

  Next, the recognition process will be described. FIG. 10 is a flowchart showing an outline of recognition processing by the pattern recognition apparatus 1. Hereinafter, the recognition processing by the pattern recognition apparatus 1 will be described with reference to FIG.

S401: A test pattern X that is a time-series pattern is input.
S402: The recognizing unit 20 calculates the likelihood C (x _i , S _j ) from the learning result of the SOINN 12 stored in the template model learning result DB 32, and performs DP matching between the template model of each class and the test pattern X. .
S403: Output the attribution class of the template model having the maximum likelihood Q (i, j). Thereby, for example, when the likelihood with respect to the template model of class 3 is maximized, the test pattern X is recognized as class 3.

Then, the effect by the pattern recognition apparatus 1 which concerns on Embodiment 1 of this invention is demonstrated. In the following, a method that models the pattern recognition apparatus 1 is referred to as a SOIN-DP method.
In order to confirm the effectiveness of the SOIN-DP method, a verification experiment was performed using actual data. In the experiment, two types of data sets of phoneme data and motion data obtained from moving images were used in order to evaluate a general-purpose learning function of time series data. In addition, comparative experiments with the conventional method HMM and stochastic DP method were conducted. Before explaining the verification experiments and results by each method, the HMM and the stochastic DP method, which are conventional methods, will be briefly described first, and then the parameter setting for the SOIN-DP method will be described.

First, a conventional method, Hidden Markov Model (HMM), will be briefly described. The HMM is an effective statistical method for modeling uncertain time-series data, and is defined as a nondeterministic probability finite automaton in the sense that a state transition destination is not uniquely determined by an output symbol. Here, the HMM parameters include three parameters: a state transition probability, a symbol output probability, and an initial state probability.
HMMs are classified into discrete HMMs and continuous distribution HMMs according to the symbol output probability calculation method. In speech recognition / motion recognition, a continuous distribution HMM is generally used. Therefore, in the first embodiment, a continuous distribution HMM is employed as a comparison method.
In addition, HMMs are classified according to topology (state connection relation) into an all-transition type (Ergodic) model capable of transitioning from one state to all states, and a left to right model in which state transition proceeds in a certain direction. The Since the left to right model is generally used in the field of speech recognition and motion recognition, the left to right model is employed as a comparison method in the first embodiment.
A generally used Baum Welch algorithm was used for the parameter estimation method of the HMM. Further, in order to improve the parameter estimation accuracy by the Baum Welch algorithm, the Segmental kmans method is used for setting the initial value of the parameter.

Next, the conventional stochastic DP method will be briefly described.
The recurrence formula used in the stochastic DP method is shown below. The recurrence formula is configured on the basis of the asymmetric recurrence formula shown in Equation 4 above.

The condition probability P (a _i | j) and the state transition probability P _DP1,2,3 (j) in the recurrence formula shown in Equation 16 above were calculated by the method described in Non-Patent Document 4. Here, the conditional probability P (a _i | j) is a multidimensional normal distribution. Regarding the covariance matrix of P (a _i | j), the same one was used in a certain range. For example, when the same covariance matrix is used for 10 states, one covariance matrix is calculated by maximum likelihood estimation from all data distributed to states 1 to 10, and states 1 to 10 The same σ is used in each state. In the states 11 to 20 and the states 21 to 30, calculation is performed by the same operation, and the same σ is used for these states.

  Subsequently, parameter setting related to the SOINN-DP method will be described. Preliminary experiments were performed using isolated words, and parameters of the SONN-DP method were set. The experiment used data in which three male speakers uttered 5 words 50 times each (150 per word, 750 in total). Word is a 5 words of "Good evening", "Hello", "Tomorrow", "Good morning", "good-bye". As the audio feature amount, a 20-dimensional MFCC, a frame length of 25 ms, and a frame period of 5 ms were used. For each class, the training data was 50 and the test data was 100, and a total of 20 cross-validation experiments were performed while exchanging the training data and the test data. From 20 cross-validation experiments, the recognition rate for the test data in each experiment was determined, and the average value was used as the recognition result. A parameter having the maximum average recognition rate with respect to the test data was used for a phoneme recognition experiment and a motion recognition experiment described later.

First, parameter setting of the SOINN 12 in the SOINN-DP method will be described.
Like SOIN 12, in SOIN 12, it is necessary to set two parameters, λ and age _dead , in order to appropriately remove noise. Each cluster is represented by a side connecting the nodes, and the node and the side are repeatedly generated and deleted according to input data. For this reason, when a node is generated for noise data, although the classification result is deteriorated, this can be avoided by appropriately setting the parameters λ and age _dead .

  λ indicates a period for deleting noise and a strange node. Although it is possible to frequently perform noise processing by setting λ to a small value, if it is extremely small, a node that is not actually noise is erroneously deleted. On the other hand, when λ is extremely increased, a node generated due to the influence of noise cannot be appropriately removed. Therefore, from the result of the preliminary experiment for parameter setting, the number of inputs to the SOINN 12 was set to 30000 times and λ = 10000. In other words, during learning, noise and ghost nodes were deleted three times.

The age _dead is used to delete an edge that is generated erroneously due to the influence of noise or the like. By setting age _dead to a small value, edges can be easily deleted and the influence of noise can be reduced. However, when the value is extremely small, edges are frequently deleted and the learning result becomes unstable. On the other hand, when age _dead is made extremely large, the side generated due to the influence of noise cannot be removed appropriately. In the SOIN-DP method, since the number of element vectors distributed to one state is small, the learning result becomes unstable when the age _dead is reduced. Therefore, in the first embodiment, the age _dead is not allowed to function and the sides are not deleted. That is, it is assumed that age _dead = 30000 and no edge is deleted during learning.

As described above, from the result of the preliminary experiment, the parameter setting of the SOINN 12 was set as λ = 10000 and age _dead = 30000. Although there are parameters c ₁ , α ₁ , α ₂ , α ₃ , β, γ in addition to the parameter λ and age _dead , the same values as those of SOIN 12 are used for these parameters (c _{1 = 1, α 1 = 1} /4, α 2 = 3/4, α 3 = 1/4, β = 2/3, γ = 3/4).

Next, parameter setting of SOINN-DP will be described.
Here, a setting method for the segment number L shown in the above equation 15 will be described. When the number of segments L is increased, the amount of input data to the SOINN 12 in each state increases, and it is considered that the learning accuracy of the SOINN 12 is improved. However, when the number of segments L is set to an extremely large value, data separated in time series is input to one state, and time series data is ignored. For this reason, excessive time-series data features cannot be modeled, leading to a reduction in the recognition rate for test data. On the other hand, when the number of segments L is set to an extremely small value, a network (a set of nodes and sides) is not formed in the SOINN 12. When a node set is not generated, it is difficult to accurately calculate the covariance matrix Σ shown in Equation 10 above. That is, in order to calculate the covariance matrix Σ shown in Equation 10, it is necessary to input a sufficient amount of data for one state (SOINN 12). In Non-Patent Document 4, the number of data is required to be at least p × 4 to 5 times the dimension number p of the feature quantity, and p ² or more is preferable.

Here, when N pieces of training data are used for model learning, it is assumed that the number of element vectors distributed to each state is N on average. In such a case, the number of elements (data number DN) of the data set Z ^* _i input to one SOINN 12 is defined by the following equation.
Therefore, from the above equation 17, the segment number L falls within the range shown in the following equation.
Through the preliminary experiment for parameter setting, the optimum number of segments L within the range of Equation 18 was determined as the minimum value satisfying L ≧ 6 p / N.
Further, since the number of segments L is considered to correspond to the range sharing the covariance matrix in the stochastic DP method, a similar preliminary experiment was performed using the stochastic DP method. As a result, in the stochastic DP method, it was confirmed that the range in which the maximum recognition rate can be obtained for the range sharing the covariance matrix is equal to the number L of segments.
Therefore, in the following experiment, the range sharing the covariance matrix in the stochastic DP method is L. The contribution of the segment number L to the recognition rate in the SOINN-DP method will be described later.

  Subsequently, in order to confirm the effectiveness of the SOIN-DP method, a recognition experiment was conducted on English phonemes as phoneme data. English text contained in the med-TIMIT database (University of Edinburgh, Center for Speech Technology Research, "CSTR US KED TIMIT" (2002), http://festvox.org/dbs/dbs_kdt.html.) A total of 3,900 phoneme 39 classes extracted from the above were used. FIG. 11 is a table showing details in the recognition target experiment. As for the parameters of the SOINN-DP, the number of segments L is determined from the above equation 18, L = 6 when the number of training data is 40, and L = 3 when the number of training data is 80.

In comparison with the conventional method, an experiment was performed under the same conditions as shown in FIG. The output probability of each state of the HMM is a mixed normal distribution having a total covariance matrix. As for the HMM, the parameters that can obtain the maximum recognition rate (number of states and number of mixtures of the mixed normal distribution) were searched for the optimum parameters by performing experiments while changing those parameters. Then, the recognition rate was calculated when such optimal parameters were used, and the result was recognized by the HMM.
For the stochastic DP method, in addition to the case of using the asymmetric recurrence formula shown in the above equation 16, an experiment using the symmetric recurrence formula was also conducted. This is for comparing the stochastic DP method using the same symmetric recursion formula with the SOIN-DP method using the symmetric recursion formula. As the symmetric recurrence formula, a formula obtained by exchanging C () shown in Equation 5 above with a conditional probability P (a _i | j) was used. Also, the range sharing the covariance matrix is between 6 states when the training data is 40, and between 3 states when the training data is 80, like the number of segments L in SOINN-DP. It was.

  FIG. 12 is a table showing average recognition results for test data obtained from the results of 10 cross-validation experiments. In FIG. 12, the first row shows the average recognition rate when the number of training data is 40 (TD40), and the second row shows the average recognition rate when the number of training data is 80 (TD80). “SO-DP” uses the SOIN-DP method, “ST-DP (1)” uses the asymmetrical recursion equation, and “ST-DP (2)” uses the symmetric recursion equation. Each of the stochastic DP methods is shown. In the right parenthesis of the recognition rate of the HMM, parameters (S: number of states, M: number of mixtures) when the maximum recognition rate is obtained are shown. As shown in FIG. 12, the average recognition rate of the SOINN-DP method was higher than the average recognition rate by the stochastic DP method and the HMM, which was good.

Moreover, in the experiment using the HMM, the number of states was varied from 1 to 13 for comparison, and as a result, the recognition rate was the maximum in the vicinity of the number of states of 3 to 7, so it was used for the experiment. The optimal number of states for phoneme data was estimated to be 3-7. Therefore, when the number of states was 3 to 7, the output probability assigned to each state was changed to a mixed continuous probability distribution, and the experiment was performed under the conditions shown in FIG. As a result, the recognition rate was maximum in the case of 40 training data in the 5 state 2 mixture, and in the case of 80 training data, the recognition rate was the maximum in the 5 state 4 mixture.
Furthermore, for the stochastic DP method, a higher recognition result was obtained when the symmetric recurrence formula was used than when the asymmetric recurrence formula disclosed in Non-Patent Document 4 was used. This is thought to be because it becomes easier to absorb time-series expansion and contraction by using a symmetric recurrence formula.
From the above, as a result of the speech recognition experiment, the SOIN-DP method was able to obtain a better recognition rate than the maximum recognition rate obtained by the stochastic DP method and the HMM. That is, the SOIN-DP method has higher recognition accuracy than the conventional method.

  Subsequently, in order to confirm the effectiveness of the SOIN-DP method, a recognition experiment was performed on motion data obtained from a moving image. As motion data, seven types of whole body motions (motions) by humans taken directly from a monocular camera were used. FIG. 13 is an image showing seven types of operation contents (operations M1 to M7) used in the experiment. The frame rate of the moving image is 29 frames per second, and the time length of each operation is a minimum of 110 frames and a maximum of 440 frames. The input pattern includes individual differences and includes elasticity in each part of the operation.

In the experiment, the features of motion were directly acquired from moving images without using a device such as motion capture. Here, in the first embodiment, a local autocorrelation feature (Nobuyuki Otsu, “Mathematical research on feature extraction in pattern recognition”, vol. 818, (1981)), which is a position invariant feature, is used for learning. Thus, dynamic features were extracted. The moving image processing procedure used in the experiment is shown below.
S501: The input moving image is smoothed and a difference between frames is acquired.
S502: The RGB value of the difference image is converted into a luminance value, and binarized from a threshold value for the luminance value.
S503: Extract autocorrelation features in the time direction between difference images. Non-Patent Literature (T. Kobayashi and N. Otsu, "Action and simultaneous multiple persons identification using cubic higher-order local auto-correlation", Proc. International Conference on Pattern Recognition, Vol. 19, pp. 741-744 (2004) For the autocorrelation feature disclosed in (.), Here, the autocorrelation feature is extracted using a 3 × 3 size mask, and only the time-series direction between the frames is extracted. The mask value at the center position is excluded because the directional feature of “movement” does not appear. As a result, a total of eight-dimensional input vectors (element vectors) were obtained in each frame.

FIG. 14 is a table showing details in the recognition target experiment. The same parameters as in the phoneme recognition experiment were used for the parameters in the SOINN-DP. However, the number of segments L was calculated as L = 5 because the input dimension is 8 when the number of training data is 10 from Equation 18.
The output probability of each state of the HMM is a normal distribution having a total covariance matrix. Similar to the phoneme recognition experiment, the experiment was performed while changing the number of states, the optimum parameter was searched, and the recognition rate using that parameter was used as the recognition result.
For the stochastic DP method, in addition to the case of using the asymmetric recurrence formula shown in the above equation 16, an experiment was also conducted in the case of using the symmetric recurrence formula. In the stochastic DP method, the range in which the covariance matrix is shared is between five states.

FIG. 15 is a table showing recognition results for the operation data. In FIG. 15, “SO-DP” is the SOIN-DP method, “ST-DP (1)” is the stochastic DP method using the asymmetric recurrence formula, and “ST-DP (2)” is the symmetric recurrence method. The stochastic DP method using an equation is shown. In the right parenthesis of the HMM recognition rate, a parameter (S: number of states) when the maximum recognition rate is obtained is shown. As shown in FIG. 15, as in the case of the phoneme recognition experiment, the discrimination rate of the SOIN-DP method is higher than the recognition rate by the stochastic DP method and the HMM, and is good.
Further, in the experiment using the HMM, the number of states was varied from 1 to 15 for comparison, and as a result, the recognition rate was the maximum at 11 states.
Furthermore, for the stochastic DP method, as in the case of the phoneme recognition experiment, the case of using the symmetric type recursion formula is higher than the case of using the asymmetric type recurrence formula disclosed in Non-Patent Document 4. It became a recognition result.
As described above, the SOIN-DP method has high recognition accuracy even for motion data in which it is difficult to determine the number of states and its output distribution compared to phoneme data.

  As described above, as in the case of the phoneme recognition experiment, the motion recognition experiment can obtain a better recognition rate than the maximum recognition rate obtained by the stochastic DP method and the HMM in the SOIN-DP method. did it. That is, the SOIN-DP method can be used for recognition of the entire time series pattern and has higher recognition accuracy than the conventional method.

Here, from the results of the phoneme recognition experiment and the motion recognition experiment, the point that the SOIN-DP method is superior to the conventional stochastic DP method and HMM will be further described.
First, the SOINN-DP method constructs a robust template model by approximating one state in detail using the SOINN 12. Thereby, the SOIN-DP method has an excellent recognition rate as compared with the stochastic DP method, and it is possible to perform robust modeling of time-series data.
In the SOINN-DP method, the output distribution in each state can be automatically determined by the SOINN 12. Moreover, since the number of states can be determined as the number of time series of the standard pattern, it is not necessary to determine the number of states in advance. As a result, when learning time-series data, the HMM needs to determine the number of states and the output distribution of states in advance (the number of mixtures in the case of a continuous distribution), but is not necessary with the SOIN-DP method. That is, in the experiment, values for the highest recognition rate were adopted for the number of HMM states and the number of mixtures, and recognition results were obtained based on these values. The SOINN-DP method was better than the recognition result by such an HMM. Therefore, the SOIN-DP method can obtain a high recognition rate without setting the number of states and the output distribution parameters in advance.

Here, the contribution of the number of segments L to the recognition rate of the SOINN-DP method will be described. The number of segments L, which is a parameter of the SOINN-DP method, affects the recognition performance of the SOINN-DP method. Therefore, the influence on the recognition accuracy due to the change in the number of segments is verified.
In the first embodiment, the number of segments L is calculated based on the above equation 18. For the verification, a phoneme recognition experiment was performed under the same conditions as in the table shown in FIG. The number of training data per class used for one experiment was 40. In addition, when the number of segments was changed from 1 to 10 and each segment number was used, a total of 10 cross-validation experiments were performed.
FIG. 16 is a diagram illustrating a verification result. As shown in FIG. 16, the recognition rate gradually increased as the number of segments was increased from L = 1, and the maximum recognition rate (57.04%) was obtained at L = 5. When the number of segments was further increased, the recognition rate declined.
On the other hand, the number of segments calculated from the results of the preliminary experiment is L = 6 when the number of training patterns is 40, and in FIG. 16, when L = 6, the recognition rate is the third highest overall. Met. Therefore, the method for estimating the number of segments used in the first embodiment is appropriate, and an appropriate number of segments can be determined by this estimation method.

Here, contribution by the SOINN 12 to the recognition performance of the SOINN-DP method will be described.
In the SOIN-DP method, the element vector group Z ^* _j is distributed to each state j by DP matching. Distributed ^Z _{* j} are classified by SOINN12, global likelihood _P whole than the classification result _(x i | _{S j)} and local likelihood _P class _| calculates _(x i _{U jk),} these probabilities A likelihood C (x _i , S _j ) is calculated based on the value. Therefore, in addition to the SOINN-DP method, two methods that do not use the SOINN 12 are defined, and these methods are compared. By comparing with the method that does not use the learning result by SOINN12, the contribution to the recognition accuracy by SOINN12 was verified. The following two methods were defined as comparison methods that did not use the learning results of SOINN12.

Method 1: The element vector group Z ^* _j is not input to the SOIN 12, and a multidimensional normal distribution P (x | S _j ) is calculated directly from Z ^* _j by maximum likelihood estimation. The likelihood C (x _i , S _j ) = log (P (x | S _j )) is assumed, and the input data is recognized by a recurrence formula using this likelihood C (x _i , S _j ).
Method 2: an element vector group ^Z _{* j} input to SOINN12, from classification result _{SOINN12, P} whole | was calculated _(x i _{S j).} However, P _class (x _i | U _jk ) obtained from the clustering result of SOINN 12 is not used for input pattern recognition. That is, the likelihood C (x _i , S _j ) is calculated based on the following formula, and α = 0.
The verification experiment was performed under the same conditions as in the table shown in FIG. 11. The number of training data per class used for one experiment was 80, and the number of segments L was 3. A total of 10 cross-validation experiments were performed for each α, while changing α from 0 to 1.0 by 0.05.

  FIG. 17 is a table showing the results of verification experiments obtained by [Method 1], [Method 2], and the SOIN-DP method. As shown in FIG. 17, the recognition result by the SOINN-DP method exceeds the recognition rate by [Method 1] and [Method 2] by about 4%. Therefore, in addition to using SOINN12, the SOINN-DP method using the information of the inner class of the learning result of SOINN12 is higher than [Method 1] and [Method 2] which did not use this information. It has a recognition rate. Note that the recognition rates of [Method 1] and [Method 2] for the test data were almost the same. That is, the SOINN-DP can further improve the recognition rate by using the information of the inner class that is the learning result of the SOINN 12.

Further, regarding the inner class information, the contribution of the change in α to the recognition rate of the SOINN-DP will be described. FIG. 18 is a diagram illustrating a change in recognition rate with respect to a change in α. In FIG. 18, the x-axis direction indicates the value of α, and the y-axis direction indicates the recognition rate corresponding to each α value. As shown in FIG. 18, the recognition rate reached its maximum at α = 0.45 and decreased thereafter. Eventually, the recognition rate was lowest at α = 1.0. The state of α = 1.0 is equivalent to calculating the likelihood only by the second term (log (Σ _j P _cass (x))) on the right side in the above equation 19, and uses only the information of the inner class. Thus, the likelihood is calculated. In this case, since each inner class is approximated by a kernel function, the correlation between dimensions cannot be modeled like a multidimensional normal distribution. For this reason, when the likelihood is calculated using only the information of each internal class, it is considered that the recognition rate for the test data is lowered.
On the other hand, from the results shown in FIG. 17, the SOIN-DP method was able to improve the recognition rate by using global information (all nodes of SOINN) in addition to the information by each internal class.
Further, as shown in FIG. 18, the maximum recognition rate was obtained at α = 0.45. This indicates that it is possible to further improve the recognition accuracy by the SOIN-DP method by fitting α to data.

In the first embodiment, the pattern recognition device 1 is described as performing the template model generation process and the recognition process. However, the present invention is not limited to this. For example, the template model generation unit 10 can be used as a template model generation apparatus that generates a template model using a self-propagating neural network. The recognition unit 20 can also be used as a recognition device that performs recognition using a self-propagating neural network.
Moreover, the pattern recognition shown in this Embodiment 1 is not limited to this, and SOIN-DP can be applied also to recognition of other time series patterns in a real environment.
Furthermore, the SOIN-DP can be applied to an intelligent robot that operates in a real environment. By applying the SOIN-DP, it is possible to make the intelligent robot recognize an operation such as sign language. In addition, by sequentially updating the template model, it can be developed to adapt to the environment. That is, for example, the cognitive development function can be realized in the same manner as a human being can gradually hear English.

  In the first embodiment, a case has been described in which a sufficient number of training patterns are prepared in advance for template model generation and a template model is generated as batch learning. However, the present invention is not limited to this. That is, when the template model generation unit 10 sequentially adds and inputs input patterns, the template model corresponding to the class to which the sequentially added input pattern belongs, the state output distribution in the template model May be updated according to the input pattern added sequentially. As a result, additional input patterns can be easily learned, and a large amount of data is not required in advance, and recognition accuracy is improved based on a small amount of data given sequentially. be able to. That is, after a template model is generated using a small amount of training patterns, the template model can be updated using supervised data provided online.

  FIG. 19 is a diagram for explaining how the template model is updated. As shown in FIG. 19, after data 1 to 3 additionally provided online are matched by the SOIN-DP method in the same manner as described above, a template model (Target Template) to be updated from supervised data is obtained. ) Can be specified. Then, each state (SOINN-1 Frame to SOIN-3 Frame) in the template model is updated with the data before the update and the added data. FIG. 20 is a diagram for explaining a state of update in the SINN 12. As shown in FIG. 20, in addition to the nodes that already exist in the SOINN 12, a class (Class generated by update data) composed of additionally input nodes is generated without destroying existing knowledge. FIG. 21 is a table showing a verification result by online supervised learning. As shown in FIG. 21, when a template model is first generated using a predetermined number (10, 20, 40) of data in batch learning and the template model is updated by adding data online, the recognition rate The improvement result of was verified. From the verification results, it can be seen that the recognition rate improves as data is added online. That is, the time series model can be robustly updated by online supervised learning.

Other Embodiments of the Invention
As mentioned above, although this invention was demonstrated by the embodiment, this invention can be variously deformed in the range of the meaning. For example, instead of SOINN 12, Enhanced-SOINN (hereinafter referred to as E-SONN) based on SOINN may be used.

  E-SOIN can separate a high-density overlapping class in the distribution of input patterns compared to SOIN. In addition, in the process of detecting the overlapping region of the distribution, since a smoothing method is introduced, the operation can be performed more stably than the SOIN. Furthermore, noise nodes can be deleted efficiently even with a single-layer structure. Furthermore, since the operation is performed with fewer parameters compared to SOIN, the processing can be executed more easily.

  The E-SOINN will be briefly described below. FIG. 22 is a flowchart showing a processing outline of learning processing by E-SOINN. The description of the same processing as that of the above-described SINN 12 is omitted. First, in S601 to S605 shown in FIG. 22, the same processing as that of SOINN 12 shown in FIG. 5 is performed. Therefore, the processing from S606 shown in FIG. 22 will be described below.

S606: The edge connection determination means determines the first winner node based on the node density of the first winner node a ₁ and the second winner node a ₂ for the node, node density, and edge between the nodes stored in the temporary storage unit. It is determined whether or not an edge is connected between a ₁ and the second winner node a ₂ , and the result is stored in the temporary storage unit.
S607: As a result of the determination in S606 stored in the temporary storage unit, when generating and connecting an edge between the first winner node a ₁ and the second winner node a ₂ , the edge connection means is stored in the temporary storage unit. For the stored nodes and the sides between the nodes, the sides are connected between the first winner node and the second winner node, and the result is stored in the temporary storage unit. Then, E-SOIN calculates the age of the side and the age of the side stored in the temporary storage unit, and the age of the side for the newly generated side and, if the side has already been generated between the nodes. The result is set to 0, the result is stored in the temporary storage unit, the age of the side directly connected to the first winner node a ₁ is incremented (increased by 1), and the result is stored in the temporary storage unit.

On the other hand, the result of determination in S606 stored in the temporary storage unit, when not connected to edges between the first winning node a ₁ and the second winning node a _2, although the process proceeds to S608, already between nodes If the side has been generated, the side deletion means deletes the side between the first winner node a ₁ and the second winner node a ₂ for the node stored in the temporary storage unit and the side between the nodes, The result is stored in the temporary storage unit.
Next, for the node and node density point values stored in the temporary storage unit, for the first winner node a ₁ , the node density calculation means calculates the node density points of the first winner node a ₁ stored in the temporary storage unit. By calculating the value and storing the result in the temporary storage unit, the node density point value calculated and stored in the temporary storage unit is added to the point value calculated previously and stored in the temporary storage unit. Accumulate as density points and store the result in a temporary storage.
Next, the E-SOIN increments the cumulative number M _a1 that the first winner node a ₁ stored in the temporary storage unit becomes the first winner node (increases by 1), and stores the result in the temporary storage unit.

S608: weight vector updating means for weight vector stored nodes and nodes in the temporary storage unit, respectively the input vector the weight vector of the weight vector of the first winning node a ₁ and first winning node a ₁ of adjacent node ξ The data is updated so as to be closer to, and the result is stored in the temporary storage unit. In E-SOINN, in order to support additional learning, instead of the input vector input count t, the cumulative number M _{a1 of} the first winner node a ₁ stored in the temporary storage unit becomes the first winner node. Is used.
S609: E-SOINN, for stored sides in the temporary storage unit, delete the edges with a exceeds a preset threshold age _t stored in the temporary storage unit age, and stores the result in the temporary storage unit .

  S610: E-SOINN indicates whether the total number of input vectors ξ stored in the temporary storage unit is a multiple of λ that is preset and stored in the temporary storage unit. And the result is stored in the temporary storage unit. As a result of the determination stored in the temporary storage unit, if the total number of input vectors is not a multiple of λ, the process returns to S602 to process the next input vector ξ. On the other hand, when the total number of input vectors ξ is a multiple of λ, the following processing is executed.

S611: The distribution overlap area detection means detects the overlap area of the distribution that is the boundary of the subcluster as shown in S301 to S305 above for the subcluster and distribution overlap area stored in the temporary storage unit, The result is stored in the temporary storage unit.
S612: The node density calculation means calculates the node density points stored and accumulated in the temporary storage unit as a ratio per unit input number, stores the result in the temporary storage unit, and stores the node density of the node per unit input number. And the result is stored in the temporary storage unit.
S613: The noise node deletion unit deletes a node regarded as a noise node from the nodes stored in the temporary storage unit, and stores the result in the temporary storage unit. Note that the parameters c ₁ and c ₂ used by the noise node deletion unit in S613 are used to determine whether or not the node is regarded as noise. Normally, a node having two adjacent nodes is often not a noise, and therefore c ₁ uses a value close to 0. Further, since a node having the number of adjacent nodes of 1 is often noise, c ₂ is assumed to use a value close to 1, and these parameters are set in advance and stored in the temporary storage unit.

S614: E-SOINN is the total number of input vectors ξ stored in the temporary storage unit, and whether the total number of input vectors ξ set in advance is a multiple of LT stored in the temporary storage unit. And the result is stored in the temporary storage unit. If the total number of input vectors is not a multiple of LT as a result of the determination stored in the temporary storage unit, the process returns to S602 to process the next input vector ξ.
On the other hand, when the total number of input vectors ξ is a multiple of LT, the node stored in the temporary storage unit is output as a prototype. After completing the above processing, learning is stopped.

The node density calculation means calculates the node density of the node of interest based on the average distance between adjacent nodes for the node of interest and the node density stored in the temporary storage unit, and temporarily stores the result. Store in the department. Specifically, the node density point calculation unit calculates a node density point value p _i given to the node i based on, for example, the following expression stored in the temporary storage unit, and stores the result in the temporary storage unit: To do. The point value p _i given to the node i is given a point value calculated based on the following formula stored in the temporary storage unit when the node i becomes the first winner node. If i is not the first winner node, no points are given to node i.

Here, e _i represents an average distance from the node i to the adjacent node, is calculated based on the following expression stored in the temporary storage unit, and the result is stored in the temporary storage unit.
Incidentally, m represents the number of neighbor nodes of node i that are stored in the temporary storage unit, W _i represents the weight vector of node i that are stored in the temporary storage unit.

  Here, when the average distance to the adjacent node is large, it is considered that there are few nodes in the area including the node. Conversely, when the average distance is small, the area has many nodes. it is conceivable that. Therefore, the node density points so that a high point is awarded when the first winner node is reached in a region with many nodes, and a low point is awarded when the first winner node is found in a region with few nodes. The value calculation method is configured as described above. As a result, it is possible to estimate the density of nodes in a certain area including the nodes, so even if the nodes are located in a high-density area, the number of nodes is the first winner. Compared with SOIN where the number of times is the node density, a node density point can be calculated that has a density approximated by the input distribution density of the input vector.

The unit node density point calculation unit calculates the node density density _i per unit input number of the node i based on, for example, the following expression stored in the temporary storage unit, and stores the result in the temporary storage unit.

Here, the number of inputs of consecutively given input vectors is divided into sections for each predetermined number of inputs λ that are set in advance and stored in the temporary storage unit, and the sum of points given to node i in each section is accumulated. It is defined as point s _i . When the total number of input vectors is set to LT that is preset and stored in the temporary storage unit, LT / λ is set to the total number n of sections, and the result is stored in the temporary storage unit. The number of sections in which the total of given points is 0 or more is calculated as N, and the result is stored in the temporary storage unit (note that N and n are not necessarily the same).

The accumulated points s _i are calculated based on, for example, the following expression stored in the temporary storage unit, and the result is stored in the temporary storage unit.
Here, p _i ^{(j, k)} indicates a point given to the node i by the k-th input in the j-th section, is calculated by the node density point calculation unit described above, and the result is stored in the temporary storage unit. To do. As described above, the unit node density point calculation unit calculates the density density _i of the node i stored in the temporary storage unit as an average of the accumulated points s _i and stores the result in the temporary storage unit.

  In E-SOINN, N is used instead of n in order to support additional learning. This is because, in additional learning, points are often not given to nodes generated by previous learning, and when n is used to calculate the density, the previously learned node density gradually decreases. This is to avoid it. That is, by calculating the node density using N instead of n, even if additional learning is performed for a long time, as long as the added data is not input near the previously learned node, that node Can be maintained without changing the density. As a result, even when additional learning is performed for a long time, it is possible to prevent the node density of the node from becoming relatively small, and the input distribution of the input vector compared to the conventional method including SOINN. The node density approximated by the density can be held and calculated without being changed.

  The distribution overlap area detection means calculates a cluster, which is a set of nodes connected by the edges, with respect to the nodes stored in the temporary storage unit, the edges connecting the nodes, and the density of the nodes. Based on the node density, it is divided into sub-clusters that are a subset of the cluster, the result is stored in the temporary storage, the overlapping area of the distribution that is the boundary of the sub-cluster is detected, and the result is stored in the temporary storage .

  Furthermore, the distribution overlap area detecting means determines whether the node density is locally determined based on the node density calculated by the node density calculating means for the nodes stored in the temporary storage unit, the sides connecting the nodes, and the node density. A node search unit that searches for a node that is the largest, a first label assignment unit that assigns a label different from labels already assigned to other nodes to the searched node, and a first label assignment unit A second label assigning unit that assigns the same label as the label of the node to which the label is assigned by the first label assigning unit among the nodes to which the label is not given by the first label assigning unit When there is a direct connection between the nodes with different labels, each side is a set of nodes connected by that side. A cluster dividing unit that divides the cluster into sub-clusters that are subsets of the cluster, and when the node of interest and its adjacent nodes belong to different sub-clusters, the region including the node of interest and its adjacent nodes is A distribution overlap area detection unit that detects a distribution overlap area that is a boundary of the distribution overlap area. Specifically, for the nodes stored in the temporary storage unit, the edges connecting the nodes, and the density of the nodes, for example, the overlapping area of the distribution that is the boundary of the sub-cluster is detected as follows, and the result is Store in the temporary storage.

S701: The node search unit searches for a node having a node density that is locally maximum based on the node density calculated by the node density calculation unit for the node and the node density stored in the temporary storage unit, Store the result in the temporary storage.
S702: The first label assigning unit assigns a label different from a label already assigned to another node to the node searched in S701 for the node and the node label stored in the temporary storage unit. The result is stored in the temporary storage unit.
S703: The second label assigning unit is the node stored in the temporary storage unit, the side connecting the nodes, and the label of the node. For the node that has not been given a label by the first label assigning unit in S702, For the node connected by the side and the node to which the label is assigned to the first label giving unit, the same label as the label of the node to which the label is given by the first label giving unit is given, and the result is temporarily stored To store. That is, the same label as the adjacent node having the locally maximum density is assigned.
S704: The cluster dividing unit uses the same cluster, which is a set of nodes connected by the side stored in the temporary storage unit, for the nodes stored in the temporary storage unit, the sides connecting the nodes, and the labels of the nodes. The data is divided into sub-clusters that are a subset of the cluster composed of nodes to which labels are assigned, and the result is stored in the temporary storage unit.
S705: The distributed overlap area detection unit, when the node stored in the temporary storage unit, the side connecting the nodes, and the label of the node belong to different sub-clusters, respectively, A region including a node to be processed and its adjacent nodes is detected as an overlapping region of distribution that is a boundary of sub-clusters, and the result is stored in a temporary storage unit.

  The edge connection determination means is a first winner node when the first winner node and the second winner node are nodes located in the distribution overlap region with respect to the node, node density, and distribution overlap region stored in the temporary storage unit. Based on the node density of the second winner node, it is determined whether or not an edge is connected between the first winner node and the second winner node, and the result is stored in the temporary storage unit. Further, the edge connection determination means includes a sub-cluster determination unit that determines a sub-cluster to which the node belongs, and a vertex of the sub-cluster to which the node belongs, for the node, node density, and node sub-cluster stored in the temporary storage unit. An edge connection determination unit that determines whether to connect an edge between the first winner node and the second winner node based on the density and the density of the node.

The edge connecting means is configured to connect the edge between the first winner node and the second winner node with respect to the node stored in the temporary storage section and the edge between the nodes based on the determination result of the edge connection determination means stored in the temporary storage section. And the result is stored in the temporary storage unit.
The edge deleting means is based on the determination result of the edge connection determining means stored in the temporary storage section, and the edge between the first winner node and the second winner node with respect to the nodes stored in the temporary storage section and the edges between the nodes. And the result is stored in the temporary storage unit.
Specifically, for the nodes, node density, node sub-clusters, and edges between nodes stored in the temporary storage unit, for example, the edge connection determination means determines whether to connect edges. The edge connecting means and the edge deleting means perform edge generation and deletion processing, and store the results in the temporary storage unit.

S801: The belonging sub-cluster determining unit determines the sub-cluster to which the first winner node and the second winner node belong for each of the nodes and node sub-clusters stored in the temporary storage unit, and stores the result in the temporary storage unit. To do.
S802: If the result of determination in S801 stored in the temporary storage unit is that the first winner node and the second winner node do not belong to any subcluster, or the first winner node and the second winner node are the same subcluster The edge connecting means connects the nodes by generating an edge between the first winner node and the second winner node for the node and the edge between the nodes stored in the temporary storage unit. The result is stored in the temporary storage unit.
S803: As a result of the determination in S801 stored in the temporary storage unit, if the first winner node and the second winner node belong to different sub-clusters, the edge connection determination unit is a node stored in the temporary storage unit, For node density and edges between nodes, determine whether to connect edges between the first winner node and the second winner node based on the density of the vertices of the subcluster to which the node belongs and the density of the node, and Store the result in the temporary storage.
S804: If the result of determination by the edge connection determination unit in S803 stored in the temporary storage unit determines that it is not necessary to connect edges, the nodes stored in the temporary storage unit and the edges between the nodes are When the 1 winner node and the second winner node are not connected by an edge and the nodes are already connected by an edge, the edge deleting means is configured to store the node stored in the temporary storage unit and the edge between the nodes. The edge between the first winner node and the second winner node stored in the temporary storage unit is deleted, and the result is stored in the temporary storage unit.
S805: If it is determined that the sides need to be connected as a result of the determination by the side connection determination unit in S803 stored in the temporary storage unit, the side connection unit determines whether the node is stored in the temporary storage unit or between the nodes. Are generated between the first winner node and the second winner node, and the nodes are connected.

Here, the determination process by the edge connection determination unit will be described in detail.
First, the edge connection determination unit determines the minimum node density m among the node density density _win and the second winner node density density _sec-win of the first winner node for the node and node density stored in the temporary storage unit, for example. Calculation is performed based on the following expression stored in the temporary storage unit, and the result is stored in the temporary storage unit.

Next, with regard to the nodes stored in the temporary storage unit, the node density of the nodes, and the subclass of the nodes, the subcluster A and the subcluster B to which the first winner node and the second winner node belong respectively, The density A _max and the density B _max of the vertices of the sub-cluster B are calculated, and the result is stored in the temporary storage unit. Of the nodes included in the subcluster, the node density having the highest node density is defined as the density of the vertices of the subcluster.

Then, regarding the density A _max and B _max of the vertices of the subcluster to which the node stored in the temporary storage unit belongs, and the density m of the node, m is smaller than α _A A _max and m is smaller than α _B B _max. And the result is stored in the temporary storage unit. That is, it is determined whether or not the following inequality stored in the temporary storage unit is satisfied, and the result is stored in the temporary storage unit.

As a result of the determination, when m is smaller than α _A A _max and m is smaller than α _B B _max , the first winner node and the second winner are determined for the node and the edge between the nodes stored in the temporary storage unit. It is determined that no edge is required between the nodes, and the result is stored in the temporary storage unit.
On the other hand, as a result of the determination, if m is greater than or equal to α _A A _max or m is greater than or equal to α _B B _max , the first winner node and the first It is determined that an edge is necessary between the two winner nodes, and the result is stored in the temporary storage unit.

Thus, by comparing the minimum node density m of the first winner node and the second winner node with the average node density of the sub-cluster including the first winner node and the second winner node, respectively, the first winner node And the size of the unevenness of the node density in the region including the second winner node can be determined. That is, when the node density m between the valleys of the distribution existing between the sub-cluster A and the sub-cluster B is larger than the threshold value α _A A _max or α _B B _max , the node density shape is determined to be small unevenness. can do.

Here, α _A and α _B are calculated based on the following expressions stored in the temporary storage unit, and the results are stored in the temporary storage unit. Since α _B can be calculated in the same manner as α _A , description thereof is omitted here.
i) When A _max / mean _A −1 ≦ 1, α _A = 0.0.
ii) When 1 <A _max / mean _A −1 ≦ 2, α _A = 0.5.
iii) If 2 <A _max / mean _A -1, then α _A = 1.0.

In the case of i) where the value of A _max / mean _A is 1 or less, the values of A _max and mean _A are similar, and it is determined that the unevenness in density is due to the influence of noise. Then, the sub-cluster is integrated by setting the value of α to 0.0.
When the value of A _max / mean _A exceeds iii), it is determined that A _max is sufficiently larger than mean _A and that there are irregularities with clear density. Then, the sub-cluster is separated by setting the value of α to 1.0.
Then, in the case of ii) where the value of A _max / mean _A is other than the case described above, by setting the value of α to 0.5, the sub-clusters are integrated or separated according to the size of the density unevenness. To be.

Here, mean _A indicates the average value of the node density density _i of the node i belonging to the sub-cluster A, and N _A is calculated based on the following formula stored in the temporary storage unit, where N _A is the number of nodes belonging to the sub-cluster A. The result is stored in the temporary storage unit.

  In this way, when separating into sub-clusters, the degree of unevenness of the node density included in the sub-cluster is determined, and two sub-clusters that satisfy a certain standard are integrated into one, thereby overlapping the distribution. It is possible to prevent instability due to excessive division of sub-clusters in area detection. For example, many fine irregularities may be formed in the density distribution due to a small amount of noise and learning samples. In such a case, when the first winner node and the second winner node are located in the overlapping region of the distribution between the sub-clusters A and B, the two that satisfy certain criteria when connecting between the nodes By integrating the sub-clusters into one, smoothing can be achieved even when the density distribution includes many fine irregularities.

  The noise node deletion unit is configured to determine the node density calculated by the node density calculation unit and the adjacent node of the target node with respect to the node, the node density, the edge between the nodes, and the number of adjacent nodes stored in the temporary storage unit. Based on the number of nodes, the node of interest is deleted, and the result is stored in the temporary storage unit. Further, the noise node deletion means includes a node density comparison unit that compares the node density of the node of interest with a predetermined threshold with respect to the nodes, node density, sides between nodes, and the number of adjacent nodes stored in the temporary storage unit; An adjacent node number calculation unit that calculates the number of adjacent nodes of the node to be processed, and a noise node deletion unit that deletes the node of interest as a noise node. Specifically, for example, the nodes, the node density, the sides between the nodes, and the number of adjacent nodes stored in the temporary storage unit as described below are based on the node density and the number of adjacent nodes of the target node. The node to be deleted is deleted, and the result is stored in the temporary storage unit.

The noise node deletion means calculates the number of adjacent nodes by the adjacent node number calculation unit for the node i of interest with respect to the nodes stored in the temporary storage unit, the sides between the nodes, and the number of adjacent nodes. Store in the temporary storage. Then, the following processing is performed according to the number of adjacent nodes stored in the temporary storage unit.
i) When the number of adjacent nodes stored in the temporary storage unit is 2, the node density comparison unit compares the node density density _i of the node _i with a threshold value calculated based on the following formula stored in the temporary storage unit, for example. The result is stored in the temporary storage unit.
For the comparison result stored in the temporary storage unit, when the node density density _i is smaller than the threshold, the noise node deletion unit deletes the node for the node stored in the temporary storage unit, and the result is stored in the temporary storage unit. To store.

ii) When the number of adjacent nodes stored in the temporary storage unit is 1, the node density comparison unit compares the node density density _i of the node _i with, for example, a threshold value calculated based on the following formula stored in the temporary storage unit. The result is stored in the temporary storage unit.
If the node density density _i is smaller than the threshold for the comparison result stored in the temporary storage unit, the noise node deletion unit deletes the node for the node stored in the temporary storage unit, and temporarily stores the result. Store in the department.

iii) Regarding the number of adjacent nodes stored in the temporary storage unit, when there is no adjacent node, the noise node deletion unit deletes the node for the node stored in the temporary storage unit, and stores the result in the temporary storage unit To do.
Here, by adjusting predetermined parameters c ₁ and c ₂ that are set in advance and stored in the temporary storage unit, it is possible to adjust the behavior of noise node deletion by the noise node deletion unit.

  An object of the present invention is to supply a recording medium (or storage medium) that records a program code of software that realizes the functions of the above-described embodiments to a system or apparatus, and a computer (or CPU or MPU) of the system or apparatus. Of course, this can also be achieved by reading and executing the program code stored in the recording medium. In this case, the program code itself read from the recording medium realizes the functions of the above-described embodiment, and the recording medium on which the program code is recorded constitutes the present invention.

  Further, by executing the program code read by the computer, not only the functions of the above-described embodiments are realized, but also an operating system (OS) running on the computer based on the instruction of the program code. Of course, a case where part or all of the actual processing is performed and the functions of the above-described embodiments are realized by the processing is included.

  Further, after the program code read from the recording medium is written into a memory provided in a function expansion card inserted into the computer or a function expansion unit connected to the computer, the function expansion is performed based on the instruction of the program code. Naturally, a case where the CPU or the like provided in the card or the function expansion unit performs part or all of the actual processing and the functions of the above-described embodiments are realized by the processing is included. When the present invention is applied to the recording medium, program code corresponding to the flowchart described above is stored in the recording medium.

It is a figure which shows the functional block for implementing this invention. It is a flowchart which shows the process outline | summary of the learning process by SOIN which is a conventional method. It is a figure which shows the artificial data set input with respect to SOINN which is a conventional method. It is a figure which shows the output result of SOINN with respect to an artificial data set. It is a flowchart which shows the process outline | summary of the learning process by SOINN12 which concerns on this embodiment. It is a figure which shows the some internal class which exists in SOIN12 which concerns on this embodiment. It is a figure which shows the system configuration | structure of the pattern recognition apparatus which concerns on this embodiment. It is a flowchart which shows the process outline | summary of the template model production | generation process by a pattern recognition apparatus. It is a figure which shows making it input a training pattern into SOINN12. It is a flowchart which shows the process outline | summary of the recognition process by a pattern recognition apparatus. It is a table | surface which shows the experimental condition in a phoneme recognition experiment. It is a table | surface which shows the experimental result in a phoneme recognition experiment. It is an image which shows the operation | movement content in operation | movement experiment. It is a table | surface which shows the experimental condition in an action recognition experiment. It is a table | surface which shows the experimental result in a motion recognition experiment. It is a figure which shows the change of the recognition rate with respect to the number of segments. It is a table | surface which shows the comparison result with the method which does not use SOINN12. It is a figure which shows the change of the recognition rate by the reflection condition of global likelihood and local likelihood. It is a figure for demonstrating a mode that a template model is updated. It is a figure for demonstrating the mode of the update in SOINN12. It is a table | surface which shows the verification result by online supervised learning. It is a flowchart which shows the process outline | summary of the learning process by E-SOIN which is a conventional method.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Pattern recognition apparatus 10 Template model production | generation part 11 Standard pattern selection part 12 SOIN
121 Inter-class node insertion unit 122 Edge deletion unit 123 Node deletion unit 124 Weight vector update unit 20 Recognition unit 21 Likelihood calculation unit 211 Global likelihood calculation unit 212 Local likelihood calculation unit 31 Training pattern DB
32 Template model learning result DB
40 Computer 41 CPU
42 ROM
43 RAM
44 bus 45 input / output interface 46 input unit 47 output unit 48 storage unit 49 communication unit 50 drive 501 magnetic disk 502 optical disk 503 flexible disk 504 semiconductor memory

Claims (27)

A part or all of the input pattern is sequentially input to the neural network as an input vector, and the number of nodes is automatically increased based on the input vector and nodes described by multidimensional vectors arranged in the neural network. Using a self-propagating neural network,
A template model generation unit that generates a template model according to the feature amount of the input pattern;
A pattern recognition apparatus comprising: a recognition unit that recognizes an input pattern by matching the generated template model and the input pattern. The recognition unit
A likelihood calculating unit for calculating an output distribution of states in the template model based on the nodes;
The pattern recognition apparatus according to claim 1, wherein the degree of coincidence between the template model and the input pattern is calculated using the calculated output distribution of the state in the template model. The likelihood calculating unit
Based on all nodes arranged in the self-propagating neural network, a global likelihood calculating unit that calculates a global likelihood,
The pattern recognition apparatus according to claim 2, wherein an output distribution of states in the template model is calculated from the global likelihood. The likelihood calculating unit
For a cluster composed of nodes connected by edges, a local likelihood calculation unit that calculates a local likelihood based on nodes belonging to the cluster,
The pattern recognition apparatus according to claim 2, wherein an output distribution of states in the template model is calculated from the local likelihood. The likelihood calculating unit
Based on all nodes arranged in the self-propagating neural network, a global likelihood calculating unit for calculating a global likelihood;
A local likelihood calculator that calculates a local likelihood based on a node belonging to the cluster for a cluster composed of nodes connected by edges;
The pattern recognition apparatus according to claim 2, wherein an output distribution of states in the template model is calculated from the global likelihood and / or the local likelihood. The template model generation unit
A standard pattern selection unit that selects a standard pattern of a class to which the input pattern belongs by matching between the input patterns;
The size of the input pattern is normalized to the size of the standard pattern, and a part or all of each input pattern corresponding to each frame of the standard pattern is made to correspond to each state in the template model. The pattern recognition apparatus according to claim 1. The template model generation unit
For a set of elements consisting of a part or all of each input pattern corresponding to each state in the template model, the element set based on the number of input patterns of the input pattern and the number of dimensions of the feature quantity of the input pattern The pattern recognition apparatus according to claim 6, wherein a self-propagating neural network is input. The template model generation unit
When the input pattern is sequentially added and input, for the template model corresponding to the class to which the sequentially added input pattern belongs, the output distribution of the state in the template model is added sequentially. The pattern recognition apparatus according to claim 1, wherein the pattern recognition apparatus is updated according to an input pattern. The pattern recognition apparatus according to claim 1, wherein the matching process is performed using a DP matching method. The self-propagating neural network is
When an edge is connected between a node having a weight vector closest to the input vector and a node having a second closest weight vector,
Class that inserts the input vector as a node based on the similarity threshold of the node of interest calculated based on the distance between the node of interest and another node, and the distance between the input vector and the node of interest Internode insertion part,
A weight vector updating unit that updates a weight vector corresponding to a node closest to the input vector and a weight vector corresponding to a node directly connected to the node by an edge so as to be closer to the input vector. The pattern recognition apparatus according to claim 1. The inter-class node insertion unit is
A node having a weight vector closest to the input vector to be inputted is a first winner node, a node having a weight vector closest to the second is a second winner node, and between the first winner node and the second winner node When the side is connected to
When there is a node that is directly connected to the target node by an edge with respect to the target node, among the directly connected nodes, the node having the maximum distance from the target node When the distance is the similarity threshold and there is no node directly connected to the node of interest by the side, the distance between the nodes having the smallest distance from the node of interest is the similarity threshold. A similarity threshold calculation unit to calculate,
Whether the distance between the input vector and the first winner node is greater than the similarity threshold of the first winner node, and the distance between the input vector and the second winner node is similar to the second winner node A similarity threshold determination unit that determines whether or not it is greater than the degree threshold,
The pattern recognition apparatus according to claim 10, further comprising: a node insertion unit that inserts the input vector as a node at the same position as the input vector based on a similarity threshold determination result. The self-propagating neural network is
Based on the age of the side associated with the side, a side deletion unit that deletes the side,
The pattern recognition device according to claim 10, further comprising: a node deletion unit that deletes the node of interest based on the number of sides directly connected to the node of interest, based on the number of sides directly connected to the node of interest. . The pattern recognition apparatus according to claim 1, wherein the self-propagating neural network has a one-layer structure. A part or all of the input pattern is sequentially input to the neural network as an input vector, and the number of nodes is automatically increased based on the input vector and nodes described by multidimensional vectors arranged in the neural network. Using a self-propagating neural network,
A template model generation step for generating a template model according to the feature amount of the input pattern;
A pattern recognition method comprising: a recognition step of recognizing the input pattern by matching the generated template model and the input pattern. The recognition step includes
A likelihood calculating step of calculating an output distribution of states in the template model based on the nodes;
The pattern recognition method according to claim 14, wherein the degree of coincidence between the template model and the input pattern is calculated using the calculated output distribution of the state in the template model. The likelihood calculating step includes:
A global likelihood calculating step for calculating a global likelihood based on all nodes arranged in the self-propagating neural network;
The pattern recognition method according to claim 15, wherein an output distribution of states in the template model is calculated from the global likelihood. The likelihood calculating step includes:
For a cluster composed of nodes connected by edges, a local likelihood calculating step for calculating a local likelihood based on nodes belonging to the cluster,
The pattern recognition method according to claim 15, wherein an output distribution of states in the template model is calculated from the local likelihood. The likelihood calculating step includes:
A global likelihood calculating step for calculating a global likelihood based on all nodes arranged in the self-propagating neural network;
A local likelihood calculating step for calculating a local likelihood based on a node belonging to the cluster for a cluster composed of nodes connected by edges;
The pattern recognition method according to claim 15, wherein an output distribution of states in the template model is calculated from the global likelihood and / or the local likelihood. The template model generation step includes
A standard pattern selection step of selecting a standard pattern of a class to which the input pattern belongs by matching between the input patterns;
The size of the input pattern is normalized to the size of the standard pattern, and a part or all of each of the input patterns corresponding to each frame of the standard pattern is made to correspond to each state in the template model. The pattern recognition method according to claim 14. In the template model generation step,
For a set of elements consisting of part or all of each input pattern corresponding to each state in the template model, the element set based on the number of input patterns of the input pattern and the number of dimensions of the feature quantity of the input pattern The pattern recognition method according to claim 19, wherein the pattern recognition method is input to the self-propagating neural network. In the template model generation step,
When the input pattern is sequentially added and input, the output distribution of the state in the template model is sequentially added to the template model corresponding to the class to which the input pattern to be sequentially added belongs. 15. The pattern recognition method according to claim 14, wherein updating is performed according to the input pattern. The pattern recognition method according to claim 14, wherein the matching process is performed using a DP matching method. The self-propagating neural network is
When an edge is connected between a node having the closest weight vector to the input vector and a node having the second closest weight vector,
Class that inserts the input vector as a node based on the similarity threshold of the node of interest calculated based on the distance between the node of interest and another node, and the distance between the input vector and the node of interest Inter-node insertion step;
A weight vector update step of updating a weight vector corresponding to a node closest to the input vector and a weight vector corresponding to a node directly connected to the node by an edge so as to be closer to the input vector. The pattern recognition method according to claim 14. The inter-class node insertion step includes:
A node having a weight vector closest to the input vector to be inputted is a first winner node, a node having a weight vector closest to the second is a second winner node, and between the first winner node and the second winner node When the side is connected to
When there is a node that is directly connected to the node of interest by a side with respect to the node of interest, among the nodes that are directly connected, between the nodes having the maximum distance from the node of interest When the distance is the similarity threshold and there is no node directly connected to the target node by the side, the distance between the nodes having the smallest distance from the target node is the similarity threshold. A similarity threshold calculation step to calculate;
Whether the distance between the input vector and the first winner node is greater than the similarity threshold of the first winner node, and the distance between the input vector and the second winner node is similar to the second winner node A similarity threshold determination step for determining whether or not it is greater than the degree threshold;
24. The pattern recognition method according to claim 23, further comprising a node insertion step of inserting the input vector as a node at the same position as the input vector based on a similarity threshold determination result. The self-propagating neural network is
A side deletion step of deleting the side based on the age of the side associated with the side;
The pattern recognition method according to claim 23, further comprising: a node deletion step of deleting the node of interest based on the number of sides directly connected to the node of interest, based on the number of sides directly connected to the node of interest. . 15. The pattern recognition method according to claim 14, wherein the self-propagating neural network has a single layer structure.   27. A program for causing a computer to execute the pattern recognition processing according to claim 14.