JP2016162437A - Pattern classification device, pattern classification method and pattern classification program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern classification device for making vector data high-dimensional based on multiple linear mapping and extracting and classifying pattern features of the vector data, a pattern classification method and a pattern classification program.SOLUTION: The pattern classification device comprises: a preprocessing part which generates an eigen vector set common for identity classes by performing main component analysis on a learning input pattern beforehand, and reconfigures an orthogonal vector from the input pattern and the eigen vector set; a high-dimensional vector conversion part which converts the orthogonal vector into a high-dimensional vector by multiple linear mapping (tensor product); a high-dimensional partial space similarity calculation part which generates out a high-dimensional eigen vector set from the high-dimensional vector for each class (k) and calculates and generates similarity for each class between the high-dimensional vector and a high-dimensional eigen vector set Ψ (k, m); and an identity classification part which performs individual identity classification from the similarity for each class.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a pattern classification device, a pattern classification method, and a pattern classification program. More specifically, the present invention relates to a pattern classification device, a pattern classification method, and a pattern classification program that have greatly improved pattern recognition performance by extracting tensor feature amounts based on feature elements of feature vectors representing input patterns.

  With the speeding up of the network and the development of a high-performance cloud computing environment, the processing for various media (text, voice, images) is changed from the main encoding body to the search body that directly accesses the content, that is, each medium. Deepening into pattern recognition / understanding has begun. In daily life, humans process patterns consisting of enormous multimodal information data input through sound / voice, text / image / video. Processing includes the ability to recognize patterns, classify patterns, and understand patterns. For example, when reading a book or a newspaper, humans understand the meaning of characters and words by collating character patterns input through vision with learned patterns. When listening to the voice, the voice pattern input through hearing is compared with the learned pattern to understand the meaning of the utterance. Furthermore, when a human watches a person's face, it is possible to identify other people or understand emotions by collating image patterns inputted through vision with learned patterns.

  Taking speech as an example, in recent years, many systems using speech, which is the most natural communication means for humans, have appeared, and speech recognition technology has also been introduced into smartphones and WEB browsers. On the other hand, with the spread of speech recognition systems, further improvement in performance of speech recognition engines is required as a basic technology.

  The current speech recognition system uses Mel-Frequency Cepstrum Coefficients (hereinafter abbreviated as MFCC) obtained from acoustic analysis as characteristic parameters, and MFCC time series as Hidden Markov Model (hereinafter HMM). The method of handling as a stochastic process is a mainstream.

  On the other hand, using a deep neural network (hereinafter abbreviated as DNN) that stacks five or more layers of multi-layer-perceptron (hereinafter abbreviated as MLP), Research on speech recognition systems that extract triphones including contexts before and after and treat these sequences as HMM probabilistic models has become active (for example, Non-Patent Document 1 and FIG. 12).

  FIG. 12 shows such a conventional voice recognition system as an example. As shown in FIG. 12, this speech recognition system includes a multi-stage MLP, that is, a DDN (used in a series of 5 to 7 stages), and when extracting phoneme features and articulatory features, A weighting factor must be set for each MLP.

  On the other hand, in the current phoneme feature extraction by DNN, MLP is extracted in several stages, so that the extraction accuracy in MLP is improved and high calculation cost is a problem (for example, Non-Patent Document 2). MLP (including DNN) has an essential problem that the convergence of the algorithm is guaranteed but the optimality is not guaranteed. Furthermore, since the internal structure of MLP and its meaning are not clear, technological progress is stagnating, and a shift to a mathematical model with a clear internal structure is desired.

Next, as a system for recognizing patterns composed of data such as voice, characters, images, and video signals, there is a system based on a subspace method (hereinafter abbreviated as SM). In this system, eigenvectors peculiar to discrete classes (phonemes, articulatory features, strokes, facial parts, etc.) at the time of learning are subspaced from vectors constituting patterns (data such as speech, characters, images, video signals). Extract as This unique eigenvector can be extracted from the learning data by principal component analysis (hereinafter abbreviated as PCA). In SM, an input pattern (vector x) whose class is unknown and M eigenvectors φ (k, m), m = 1,2, ... M, k = 1, 2, ..., K of class k The inner product is calculated, and the similarity s (k) of class k is calculated from the following equation. Where ‖ ‖ is the norm (square root of the sum of squares).

  That is, the unknown class input accumulates the inner product of the learning data with M eigenvectors to obtain the class k similarity. Here, the square of the inner product is calculated because the similarity is defined as a value that does not depend on the sign of the eigenvector. When class membership is determined, k that gives the maximum value of s (k) is used as the recognition result. Unlike MLP or DNN, SM has the advantage that convex optimization is guaranteed, but the eigenvector φ (k, m) targets only the information in the class and handles the relationship between the pattern feature classes. There is a limit to the recognition performance because there is no connection and the relationship between the elements constituting the vector is not used.

  On the other hand, there is a linear discriminant analysis (hereinafter abbreviated as LDA) as feature extraction for information between pattern feature classes. LDA is a method of identifying in consideration of the difference between classes, and discriminates a class by evaluating both within-class and outside-class covariance. However, since LDA calculates eigenvectors from surrounding intra-class covariance and extra-class covariance around the average vector for each class, the ability to separate classes increases, but it has the disadvantage of being too dependent on learning data. In addition, there is a drawback that performance degradation is large due to the difference in environment between data collection (learning) and recognition.

  Furthermore, as a system for recognizing speech data, a method of extracting phoneme discrimination feature vectors by MLP and PCA and performing phoneme recognition has been proposed (for example, Non-Patent Document 3). Also, a method has been proposed in which phoneme vectors are extracted from PCA to improve recognition performance (for example, Non-Patent Document 4).

  However, even if a feature extractor that employs any of the above methods is used, the feature is accurately extracted from a pattern (data such as voice, text, image, video signal, etc.) having an enormous amount of information, and the pattern has good performance There was a problem that could not be recognized.

  In particular, in recent years, users of SNS (Social Networking Service) such as blogs, video sites, Facebook (registered trademark), Twitter (registered trademark) have increased, and not only characters from terminals such as personal computers and smartphones, Enormous amounts of digital data such as voice, photos, and moving images are stored in various server computers on the Internet. These so-called big data are said to be several hundred trillion bytes or more. Under such circumstances, a pattern recognition device capable of identifying individual patterns with high accuracy from voice data, character data, and image data in big data is indispensable.

  However, since conventional text / sound / image processing techniques have been targeted for vector data, it has been difficult to dramatically improve the accuracy of pattern recognition. In addition, with the speeding up of the network and the development of a high-performance cloud computing environment, the processing for various media (text, voice, images) has been deepened from the main encoding body to the search body that directly accesses content. It has begun.

  Due to such changes in processing subjects, attention is focused on pattern classification from pattern recognition. Examples of pattern classification include a system that recognizes an enormous number of language patterns, and a large-scale language recognition device that has improved language recognition performance by incorporating a tensor analysis technique into a neural network has been proposed. (For example, Non-Patent Document 5).

  In order to respond to such a request, processing including the relationship between vector elements, that is, feature extraction and classification based on tensor analysis is required from processing of vector objects so far. In addition, this patent applicant presents the following publications as publications in which the above-mentioned literature known invention is described.

Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury, “Deep Neural Networks for Acoustic Modeling in Speech Recognition”, IEEE SIGNAL PROCESSING MAGAZINE [2], pp.82-97, november 2012. Mohammad Nurul Huda, Hiroaki Kawashima, and Tsuneo Nitta, “Distinctive Phonetic Feature (DPF) extraction based on MLPs and Inhibition / Enhancement Network,” IEICE Trans. Inf. & Syst., Vol.E92-D, No. 4, pp. 671-680 (2009). Fukuda, Nitta “Examination of Phoneme Feature Vector Orthogonalization Method for Robust Speech Recognition”, Information Processing Society of Japan, 2003-SLP-49, 2003. Park, Mizoguchi, Ariki “Examination of speech feature extraction by phoneme vector using PCA” Autumn Meeting of the Acoustical Society of Japan 1-p-26, 2007 B. Hutchinson, L. Deng, and D. Yu, “A deep architecture with bilinear modeling of hidden representations: applications to phonetic recognition”, Proc. ICASSP2012, pp.4805-4808, March 2012.

  In order to meet these demands, the present invention identifies patterns with high accuracy based on processing that includes the relationship between vector elements, that is, processing that includes relationships between vector elements, that is, feature extraction and class classification based on tensor analysis. A classification apparatus, a classification method, and a classification program are provided. More specifically, a classification apparatus capable of recognizing a pattern with high accuracy using a method of generating an orthogonal vector sequence combining a dual space and a tensor space is provided. More particularly, the present invention relates to a pattern classification apparatus, a pattern classification method, and a pattern classification program that can classify patterns at high speed and with high accuracy using a method of generating vector data of multiple dimensions by multi-linear mapping vector data in a tensor space.

The present invention is composed of the following technical matters.
(1) A pattern classification device capable of recognizing an input pattern by extracting a pattern feature similarity vector,
A preprocessing unit that performs preprocessing on the input pattern;
A high-dimensional vector conversion unit that converts the orthogonalized vector subjected to the preprocessing into a high-dimensional vector by multiple linear mapping (tensor product);
Generate and output a high-dimensional eigenvector set for each class from the high-dimensional vector, and calculate and generate a similarity for each class between the high-dimensional vector and the high-dimensional eigenvector set Ψ (k, m) A subspace similarity calculation unit;
A pattern classification apparatus comprising: a feature classification unit that classifies individual features based on the similarity by class.
(2) The pre-processing unit maps the input pattern to a class-common dual space by taking an inner product of the input pattern and an eigenvector set common to the feature class generated in advance by principal component analysis on the learning pattern. The pattern classification apparatus according to (1), further including a process of reconstructing an orthogonal vector common to a class by arranging the scalar values obtained from the inner product operation in the number of axes of eigenvectors.
(3) The higher-order subspace similarity calculating unit includes a competitive learning unit that corrects classification error by applying competitive learning using a mirror image to a higher-order subspace method,
Class-specific eigenvectors Ψ (k, m) obtained after competitive learning including mirror image learning between (auto) correlation matrices for each class using high-dimensional feature learning data in the competitive learning unit, (1) The pattern classification apparatus according to (1), wherein the pattern classification apparatus is used as a eigenvector set for calculating higher-order subspace similarity.
(4) The classification apparatus according to (1), further including a similarity posterior probability processing unit, wherein the similarity posterior probability processing unit receives the similarity and inputs the posterior of the class to which the input vector data belongs. A pattern classification apparatus characterized in that a probability is calculated.

  The present invention is also applied to a pattern classification method and a pattern classification program characterized by the technical matters described above.

  The present invention can extract and classify pattern features of input data with high accuracy by analyzing a relationship (tensor structure) between elements constituting an input vector based on multiple linear mapping.

  In addition, by applying competitive learning using mirror images to the subspace method in the higher-order subspace similarity calculation unit, the boundary of the hyperplane in the subspace method is corrected correctly, and the input pattern is classified more accurately. Can do.

  Further, by performing dimension compression processing of vector expression from principal component analysis of input vector data, both calculation efficiency and recognition performance in the classification device are improved.

  Further, by extracting the main component of the higher-order subspace generated from the tensor product, it is possible to effectively extract the feature quantity reflecting the correlation between the vector elements.

  In addition, the higher-order subspace extraction unit adapts competitive learning using a mirror image to the subspace method, so that the boundary of the hyperplane in the subspace method is corrected correctly, and the input pattern is classified more accurately. Can do.

  Further, when an input is given in a higher-order subspace, it can be probabilistically calculated to which class this input belongs. In other words, the posterior probability process by the similarity posterior probability processing unit is possible.

It is the model figure which showed the structure of the classification device of this invention. It is a hardware block diagram of this invention. It is explanatory drawing explaining the process of the multiple linear mapping part. It is an average pattern of feature values obtained by connecting three frames of a 24-dimensional filter bank coefficient of a phoneme and its logarithmic power. It is explanatory drawing of the average pattern of the feature-value obtained by the tensor product of the average pattern shown in FIG. It is explanatory drawing of the average pattern of the feature-value which connected 24 frame filter bank coefficient of the phoneme, and its logarithmic power 3 frames. It is explanatory drawing of the competitive learning using a mirror image. The similarity distribution P (s | p (k)) for each phoneme feature observed from the training data and the distribution shown by the similarity of all phoneme features It is a model figure of the prior distribution approximated from the distribution which can be observed in advance. Features based on multiple linear mapping extraction is processed using a classification device of the present invention is a block diagram of (in the figure v _n) and higher subspace method (likelihood s _k). It is a flowchart figure which shows the process sequence of this invention. It is a graph which shows the experimental result of the phoneme recognition concerning this invention. This is a pattern recognition device based on the phoneme feature DDN conversion of the conventional method.

  Hereinafter, embodiments of a pattern classification device, a pattern classification method, and a pattern classification program according to the present invention will be described with reference to the accompanying drawings. The attached drawings are used to explain the technical features of the present invention, and the configuration of the apparatus described, the procedure of various processes, and the like are limited to only those unless otherwise specified. Not the purpose.

  FIG. 1 is a model diagram showing a configuration of a pattern (phoneme) classification device 1. With reference to FIG. 1, the electrical configuration of the pattern classification apparatus 1 will be described. In the present embodiment, voice is used as a pattern recognition example, and the pattern classification apparatus 1 will be described as the voice recognition apparatus 1 below.

  The speech recognition apparatus 1 applies a BPF (Band Path Filter) as a preprocessing to a pattern (speech) input from a speech input device such as a microphone (not shown) to generate an acoustic feature vector sequence composed of a speech spectrum sequence. Pre-processing unit 2, orthogonal vector reconstructing unit 4 for orthogonalizing acoustic feature vector sequence using class common eigenvector set 3 common to previously learned (for example, phoneme feature) class, and tensor for orthogonal vector sequence A high-order subspace of a high-dimensional vector sequence using a multi-linear mapping unit 5 that generates a high-dimensional vector sequence by multiple linear mapping based on the basis of the product, and a class-specific eigenvector set 6 obtained in advance for each phoneme feature class from learning data High-order subspace similarity calculation unit 7 for calculating the similarity in, and similarity obtained by posterior probability from similarity series for each phoneme feature class A posterior probability processing unit 8 for converting to a degree (likelihood) sequence, an HMM acoustic model obtained by learning from a similarity sequence that has been a posteriori probability of phoneme features, and an HMM language model previously learned using language resources ( N-gram) is used by the HMM phoneme classification unit 10 for recognizing phonemes.

  The pre-processing unit 2 extracts pattern features from vector data pattern-inputted through an input unit 14 installed outside a corpus or the like, and generates a pattern feature vector series. Here, pattern input refers to vector data composed of various media (text, sound, image). A typical example of speech is a time-series vector obtained by obtaining a spectrum by Fourier transform and then passing through a bandpass filter (BPF). In text, after obtaining a word string by morphological analysis, a vector in which occurrence probabilities for each word are arranged is representative. In an image, there are a vector in which data obtained by applying image feature extraction to pixels is arranged in time and space for various objects in the image, or a vector in which occurrence probabilities of each extracted object are arranged.

  The preprocessing unit 2 has, for example, a band filter (BPF) group of about 24 channels, and the center frequency is set at mel scale intervals. The audio signal is input to a band-pass filter (BPF) group, and a spectrum pattern is output as an acoustic feature vector series.

  This acoustic feature vector sequence is the output of 25-channel mainly logarithmic spectrum including logarithmic power in the logarithmic output of BPF24 channel with the center frequency set to the Mel scale of auditory characteristics, or discrete cosine transform to logarithmic spectrum (Discrete) Mel frequency cepstrum coefficient (hereinafter abbreviated as MFCC) subjected to cosine transform (DCT) is used. In addition to extracting MFCC from a spectrum sequence by DCT, a convolution operation may be performed on a spectrum sequence, that is, a two-dimensional pattern of frequency and time, or a local feature may be extracted. .

  Note that the pre-processing unit first performs autocorrelation matrix R (j), j = 1, 2,... From acoustic feature vectors of all phoneme feature classes using principal component analysis (PCA) for these feature vectors. After obtaining J (= I × I), the eigenvalue problem of the correlation matrix is solved to obtain a class-specific eigenvector set 6, and then the class-specific eigenvector set 6 is used to share the class using the following formula: And the eigenvalue set 3 common to the class is generated by solving the eigenvalue problem of the correlation matrix common to the class. Note that vector sequence data represented by these acoustic feature vector sequences and the like are generally not orthogonalized. For this reason, it is necessary to orthogonalize the vector sequence data by the orthogonalized vector reconstruction unit 4 described below to reconstruct the vector.

The orthogonal vector reconstructing unit 4 arranges arbitrary M scalar values c (m) obtained from the inner product of the class-specific eigenvector set 3 generated in the preprocessing and the feature vector x (i), and generates the orthogonal vector y ( Reconfigure m).

  This process is called KL transformation, but is called dual transformation in linear algebra, and the function vector x (i) is converted into a vector of equivalence by arranging scalar values c (m) by a function f (φ (m)). To do. The function f (φ (m)) is regarded as a projection function based on an eigenfunction, and is called a functional because it is a function of the function.

  Through the above processing, the feature vector sequence is converted into a vector sequence having orthogonal components. Note that the original feature vector series is a vector of about 25 dimensions per frame in a pattern of one audio sample or the like, but in the orthogonalization, about 11 frames are collected together as feature vectors x (i), i = 1,. , I (= 25 × 11), c (m) is obtained by dual transformation from the eigenvector set φ (m, i) obtained as a result of the principal component analysis, and this scalar value is arranged to be orthogonalized vector y. Used as (m).

  At this time, by taking the inner product with the eigenvector corresponding to the large eigenvalue positively and ignoring the inner product with the eigenvector corresponding to the small eigenvalue, redundant components with low correlation can be removed in advance. As a result, it is possible to reduce the subsequent calculation amount while retaining the discriminating information.

A multiple-linear map unit 5 obtains a high-dimensional vector by expressing the internal structure of the orthogonal vector y (m) with a tensor. For simplicity, the case of Bi-linear Map will be described below. The multiple linear mapping is defined as a multiple linear function and an n-order tensor product expressed by the following equation on the dual space V ^{* with} respect to the vector space V.

The bilinear map is expressed as a vector formed by the following equation and a second-order tensor product.

As the vector component, the relationship between elements output by the dual space with φ (i) φ (j) as a new basis is _expressed as z _n = {y ₁ y ₁ , y ₁ y ₂ ,..., Y ₁ y _M , y ₂ y ₁ , y ₂ y ₂ ,..., y ₂ y _M ,... y _M y ₁ , y _M y ₂ ,..., y _M y _M } and M × (M + 1) / 2 quadratic forms expressed.

FIG. 3 is an explanatory diagram for explaining the processing of the multiple linear mapping unit 5. The multiple linear mapping unit 5 is a four-dimensional orthonormal vector sequence data output from the orthogonal vector reconstructing unit 4, that is, an orthonormalized vector having _four elements (x ₁ , x ₂ , x ₃ , x ₄ ). A second-order tensor product (4 × 4 table) can be created between x. The tensor product can be obtained by calculating the matrix product of a row vector x ^t is the transposed vector and column vector x. Here, when the components of the tensor product are examined in detail, for example, x ₂ x ₁ and x ₁ x ₂ , x ₃ x ₂ and x ₂ x _3, etc., the same value is obtained with the diagonal components of the same table as the boundary. There are a plurality of objects (symmetrical to each other), which are shown in white in the figure. The number of independent components (dark portions) excluding the same component is 10 by calculating as M = 4 in M (M + 1) / 2. By arranging these 10 independent components from the left column of the tensor product into one column, a high-dimensional vector g (x) that has been increased from the vector x can be generated.

  FIG. 4 shows an average pattern of feature quantities obtained by connecting three frames of a 24-dimensional filter bank coefficient of a phoneme and its logarithmic power, and FIG. 5 shows an average pattern of feature quantities obtained by their tensor product. In FIG. 5, each 3 × 3 block represents a group of three time frames before and after, and the contents of the block represent a correlation between spectra over time and frequency. In the figure, the 3 × 3 block representing the time is symmetrical with the diagonal component between the blocks, but the non-diagonal component of the 3 × 3 block has symmetry within the block. I don't have it. This is because the correlation between different frames is expressed. On the other hand, the diagonal component of the 3 × 3 block has symmetry inside the block. This is because the correlation between the same frames is expressed. In this way, the tensor feature amount calculated from a vector obtained by concatenating features based on the frequency spectrum for multiple frames is an explicit representation of the correlation between the components of each time and frequency, and a more precise spectrum on time and frequency. The structure can be extracted.

  FIG. 6 shows phonemes / b /, / d /, / g / when the spectrum series is 3 frames (left side of the figure). Each frame is displayed with 24 BPF channels. Further, on the right side of the figure, tensor sequences / b /, / d /, and / g / representing the spectrum sequences as tensors are shown. The triangular part in the tensor sequence expresses the relationship between channels in the frame (the tensor expression in the same frame is symmetric and is displayed in the triangular area). The rectangular area expresses the relationship between the different frames (1-2, 1-3, 2-3) (data is displayed as a rectangular area because of asymmetry). As an easy-to-understand example, the secondary relationship was directly displayed from the spectrum. In this embodiment, the acoustic feature vector sequence which is the spectrum data shown here is orthogonalized by the orthogonal vector reconstruction unit 4 by the class common eigenvector set 3, and bilinear mapping is performed on this orthogonalized vector sequence. . As the internal structure of the acoustic feature vector is explicitly expressed by multiple linear mapping, the dimension is greatly expanded. As a result, the difference between phoneme feature classes is expanded, and the orthogonalized vector sequence is converted to a high-dimensional vector sequence. You can see that.

Returning to the description of FIG. 1, the high-order subspace similarity calculation unit 7 calculates the similarity for each phoneme feature by the (higher-order) subspace method for the high-dimensional vector series. Specifically, an eigenvector set for each class is extracted in advance by principal component analysis (PCA) from high-dimensional vector learning data for each phoneme feature class. Then, an input pattern (high-dimensional vector x) whose class is unknown and M eigenvectors Ψ (k, m), m = 1, 2,. . The inner product of M, k = 1, 2,..., K is calculated, and the similarity s (k) of class k is calculated from the following equation.

  That is, the inner product of M eigenvectors is accumulated for an unknown class input to obtain a time series of K similarities (likelihoods). Unlike the normal subspace method, the higher-order subspace method used in the present invention is characterized in that it uses the relationship between the elements constituting the original vector sequence.

  The high-order subspace method (HSM) used in the present invention has the advantage that the convex optimization possessed by the subspace method is guaranteed, and at the same time, does not use the relationship between the elements constituting the vector sequence. The disadvantages of the conventional subspace method are also eliminated.

  The high-dimensional subspace similarity calculation unit 7 includes a competitive learning unit (not shown), and mirror image learning by the competitive learning unit (P. Perner (Ed.): MLDM 2001, LNAI 2123, pp. 239-248, 2001. c Springer-Verlag; Meng Shi, Tetsushi Wakabayashi, Wataru Ohyama, and Fumitaka Kimura; Faculty of Engineering, Mie University, 1515 Kamihama, Tsu, 514-8507, Japan. The boundary of the hyperplane is corrected correctly, and the input pattern can be classified more accurately.

The operation of mirror image learning is shown in FIG. When the pattern x belonging to the class k ′ is erroneously recognized as the class k, the mirror image of x with respect to the representative pattern y of k is y− (xy) = 2y−x. Therefore, using this y, R (k) + βyy ^T (R (k) and the autocorrelation matrix are corrected, and the subspace of class k is moved away from class k ′ (β is a small constant). The autocorrelation matrix of 'is modified in the direction to capture R (k') + αxx ^T (R (k ')) and learning pattern x with respect to R (k'). After mirror image learning, competition from R (k) The learned class-specific eigenvector set is recalculated and used in the high-dimensional subspace method, where the representative pattern y of the class k that is the pivot when obtaining the mirror image above is represented by x in the autocorrelation matrix R (k). And R (k) x (determined as y (corresponding to associative)), or is synthesized from the eigenvector Ψ (k, m) of class k with the following equation.

The high-order subspace similarity calculation unit 7 extracts a feature part included in the high-dimensional vector data output from the multiple linear mapping unit 5 by principal component analysis. That is, the higher-order subspace similarity calculation unit 7 performs principal component analysis on the high-dimensional vector data, and generates the principal components expressed therein in the form of eigenvectors. At the same time, higher subspace similarity calculation unit 7 generates likelihood s _k using subspace method from the generated eigenvectors. The higher-order subspace similarity calculation unit 7 has a competitive learning unit for correcting an erroneous classification by competitive learning using a mirror image. Here, an erroneous classification is performed by performing competitive learning using a mirror image. Corrections can be made to improve the performance of the classification device of the present invention. That is, the higher-order subspace similarity calculation unit 7 can further improve the recognition rate of speech, characters, images, and the like by applying competitive learning using a mirror image to the subspace method.

Note that the classification device of the present invention can further include a similarity posterior probability processing unit 8. The similarity posterior probability processing unit 8 performs the process of the posterior probability of using the likelihood s _k generated by the high-order subspace similarity calculator 7. The similarity posterior probability processing unit 8 performs posterior probability processing on the similarity series for each phoneme feature that is the output of the higher-order subspace similarity calculation unit 7. The posterior probabilistic is usually performed by softmax processing. Softmax

And is processed using a normalized exponential function. Here, a posterior probabilistic processing method in which a more practical equation is derived from the similarity prior distribution shown in FIG. 8A according to the Bayes equation will be described. Bayesian formula

The probability P (p (k)) that the phoneme feature is p (k) given the similarity value s is the likelihood P (s | p (k)) of the phoneme feature that is a prior distribution and It is shown that it can be calculated from the occurrence probability P (p (k)) of the phoneme feature. Since P (p (k)) is the probability of the language model, it is evaluated by the subsequent HMM phoneme classification unit. FIG. 8A shows a similarity distribution P (s | p (k)) for each phoneme feature observed from learning data and a distribution indicated by the similarity of all phoneme features. FIG. 8B shows a model of a prior distribution approximated from the distribution that can be observed in advance. From this, the prior distribution can be approximated and calculated as shown below.

The function of the state posterior probability P (k | x) is approximated as y (s _k ) = exp (as _k + b) −1. However, the symbol “exp” is the base of the natural logarithm. s _k is a likelihood, and a and b are coefficients obtained from conditional expressions described later. If the maximum value of the similarity is described as s _max and the minimum value is described as s _min , 0 = y (s _k = s _min ) and 1 = y (s _k = s _max ), so the coefficients a and b are determined. The following conditional expression is determined.

If a and b obtained from the above equation are returned to the original function y (s _k ) = exp (as _k + b) −1, the state posterior probability P (k | x) is derived as follows. However, s _max = 1 was assumed at the end of derivation.

In this way, when a certain input x is given, it can be probabilistically calculated to which class k the input x belongs.

In addition, when a comparative experiment was performed in advance with respect to the above-described _softmax , the approximate expression using _Smax , and _Smax = 1.0, there were few cases where _Smax = 1.0. Since a good result was shown, the posterior probability process when S _max = 1.0 is adopted in the phoneme recognition evaluation experiment described later.

  The feature classification unit 10 includes data of the likelihood vector S (k) generated by the higher-order subspace similarity calculation unit 7 and pattern models such as acoustic models and language models stored and stored in the feature classification unit 10. Compare The feature classifying unit 10 sends a comparison result for the likelihood vector series to the output device as a recognition result. The output result is displayed on a display (not shown) or the like, and the classification result is recognized by the user of the classification device of the present invention. For example, as the feature classification unit 10, various kinds of voice recognition software available via a communication network may be used.

  The HMM phoneme classifier 10 uses a 3-5 state probability transition network prepared for each phoneme from an acoustic model and a language model, similarly to the conventionally used HMM phoneme classifier. However, among the output probabilities and transition probabilities of the HMM acoustic model, the phoneme likelihood vector after the posterior probability process is used as the output probability. In the HMM classification unit, a phoneme model is used for a phoneme likelihood vector for an unknown input (a monophone model is used for each phoneme and a triphone model is used in consideration of a difference in phoneme context). In the HMM phoneme model, probability calculation using the output model and the transition model is performed according to the state transition, the path is selected by the magnitude of the accumulated value from the beginning of the speech, and the next is connected at the end of the phoneme model After selecting the phoneme model from the probability values of the N-gram language model (2-gram, 3-gram, ...), the search continues and the phoneme sequence that maximizes the maximum likelihood (log likelihood) at the end is The answer is correct.

  FIG. 2 shows the hardware configuration of this apparatus. The central processing unit 11 includes an input unit 14 and an input unit 15. In the pattern classification apparatus 1, the feature part extraction unit 2, the orthogonalized vector reconstruction unit 4, the multiple linear mapping unit 5, the higher-order subspace similarity calculation unit 7, the similarity posterior probability processing unit 8, and the feature classification The unit 10 is configured by the central processing unit 11 of the computer executing processing such as numerical calculation and control according to the processing procedure described below.

  In the pattern classification apparatus, a preprocessing unit 2 including BPF (bandpass filter) analysis, an orthogonalized vector reconstruction unit 4, a multiple linear mapping unit 5, a higher-order subspace similarity calculation unit 7, a similarity posterior probability generation unit 8 and the HMM phoneme classification unit 10 are configured by a central processing unit 11 of a computer executing processing such as numerical calculation and control according to a processing procedure described below. In addition, the eigenvector set 3 (common to classes), the eigenvector set 6 by class (for high-dimensional vector sequences), the HMM acoustic model, and the language model 9 are all computer storage devices that are electrically connected to the central processing unit 11. 12 is provided. In addition, the storage device 12 stores a phoneme classification program corresponding to the processing procedure executed by the central processing unit 11.

  The central processing unit 11 can use, for example, a CPU having an input / output interface. As the storage device 12, for example, a ROM (Read Only Memory), a RAM (Random Access Memory 16), an HDD (Hard Disk Drive), or the like can be used. Although not shown here, an input device such as a microphone to enable input of speaker speech, or an output device such as a display or speaker that enables output of recognition results obtained by the HMM phoneme classification unit 10, for example. May be electrically connected to the input / output interface of the central processing unit 11.

  Note that the hardware configuration of the classification device 1 in the present invention is not limited to that shown in FIG. Therefore, a part of the configuration of the classification device 1 may be electrically connected via a communication network such as the Internet.

  Moreover, although the classification apparatus 1 of the present embodiment is provided independently from other systems, the present invention is not limited to this configuration. Therefore, it is also possible to adopt a configuration incorporated as part of another device. In this case, input / output is performed indirectly via the other devices described above.

  When an input device is used, each time a large number of patterns (sounds, characters, images, video signals, faces, etc.) are input to the input device, the patterns are converted into analog electrical signals by the input device. / D converter (not shown). In response to this, the digital electric signal converted by the A / D converter is loaded into the central processing unit 11 so that the central processing unit 11 can process the captured pattern.

FIG. 9 is a block diagram of feature extraction (v _{n in the} figure) and higher-order subspace method (likelihood s _k ) based on multiple linear maps processed using the classification device of the present invention. FIG. 10 is a flowchart showing a processing procedure using the classification apparatus of the present invention. The algorithm will be described below with reference to FIGS. The algorithm generally follows the flow:

Steps S1 to S3: Principal component analysis is performed on the input data (vector) to obtain an eigenvector.
Step S4: An inner product of the eigenvector and the input vector is obtained, and a set of scalar values (= reconstructed vector) is obtained.
Step S5: The obtained vector is mapped to a tensor space, and higher order is performed in the tensor space, and a higher-dimensional vector (relationship between original vector components is explicitly expressed) is obtained.
Steps S6 to S7: A principal component analysis is performed on the obtained higher-dimensional vector to obtain an eigenvector. It learns and classifies differences in subspaces (subspaces based on eigenvectors) in which data is distributed for each class (for example, consonants).
Step S8: Post-processing by prior probability.
Step S9: Pattern data classification by the feature classification unit.

Hereinafter, the above algorithm will be described in detail in order.
Step S1 (Generation of pattern feature vector series)
A pattern feature vector sequence is generated by extracting pattern features from vector data pattern-inputted through an input unit 14 installed outside a corpus or the like. For example, when the present invention is used for speech recognition, the feature extraction unit 2 extracts the filter bank outputs x ₁ , x ₂ ,... X _T , x _t , and then, for example, outputs the filter bank output to the front and rear 11 frames. By concatenation, the following acoustic feature vectors of the phoneme feature class are obtained.

Step S2 (principal component analysis of feature vectors)
In the pre-processing unit 2, a principal component analysis (PCA) is performed on the feature vector to calculate an eigenvector Φ (k, m) for each pattern feature class.
Step S3 (Generation of pattern common eigenvector Φ (m))
By finding the autocorrelation matrix R (j), j = 1,2, ... J (= I × I) common to the pattern feature classes from the eigenvectors Φ for each pattern feature class, and solving the eigenvalue problem of the autocorrelation matrix R The common eigenvector Φ (m) is calculated.
Step S4 (reconstruction of orthogonalized vector y (m))
The scalar vector c (m) obtained from the inner product of the pattern common eigenvector Φ (m) generated in step S3 and the input feature vector x (i) is arranged to reconstruct the orthogonalized vector y (m). By this processing, the entire class is formed in a (normal) orthogonal space without redundancy.
Step S5 (higher-order feature vector by multiple linear mapping)
The reconstructed orthogonalized vector is multiple linear mapped according to the following formula.

However, x is an arbitrary d-dimensional input vector and assumes a real value. A symbol with a cross in a circle between x and x represents a direct product symbol. Therefore, the function defined by the above equation has an argument as a tensor product of the input vector x and itself, and the result is a symmetric matrix in this case. Finally, vec _tri (·) is a function that arranges the columns of the lower triangular part of the symmetric matrix and vectorizes them. However, when a matrix is not a target, for example, when using a tensor product of different input vectors to increase the dimension of a vector, the function vec _tri (·) performs an operation of arranging products of all elements of the tensor product. . That is, the function vec _tri (·) is a function that arranges independent elements of a given tensor product.
Step S6 (High-dimensional eigenvector set extraction)
An eigenvector set Ψ (k, m) for each class is extracted from the high-dimensional vector learning data for each phoneme class by principal component analysis (PCA).
Step S7 (similarity calculation)
The similarity s (k) is extracted using the inner product of the high-dimensional vector obtained in step S5 and the eigenvector set Ψ (k, m) for each class obtained in step S6. At this time, the competitive learning unit also recalculates the eigenvector set for each class.
・ Step S8 (posterior probability processing)
The degree of similarity s (k) obtained in step S7 is applied to the following formula to calculate the posterior probability. Here, s _min represents the minimum similarity value of all classes by the subspace method.

That is, the probability that the input x belongs to the class k is calculated by the similarity posterior probability processing unit.
Step S9 (pattern data is classified by the feature classifier)
Finally, the feature classification unit 10 classifies the pattern data subjected to the posterior probability processing. The feature classifying unit 10 includes a program for executing an acoustic model, a language model, and the like using a hidden Markov model in a storage device (not shown).

  The present invention can recognize and classify an input pattern through the above steps S1 to S9. Note that the classification device of the present invention can increase the accuracy of extracting features such as phonemes and articulatory features and reduce the amount of computation when the pattern (vector data of speech, characters, faces, etc.), for example, speech. And has a high speech recognition rate. The classification device according to the present invention is effective not only for speech but also for recognition using vector data such as characters, images, and faces as patterns. For this reason, the classification apparatus of the present invention recognizes and recognizes the characteristics of patterns included in big data composed of enormous data composed of voice data, character data, image data, etc., and classifies (classification) ) And discrimination.

  FIG. 11 shows an experimental result of phoneme recognition according to the present invention. The DNN / HMM in FIG. 11 identifies a phoneme (triphone) by DNN and inputs the obtained phoneme vector sequence (phoneme likelihood) to the HMM. The language model indicates when bi-gram is used and when tri-gram is used. In contrast, SM / HMM is the result of combining the conventional subspace method (SM) and HMM, and HSM / HMM is the result of combining the higher order subspace method (HSM) and HMM according to the present invention. . The above experiment was conducted using an English speech corpus (TIMIT), which is widely used for speech recognition. Among them, the results of the TIMIT core test used by many research institutes were shown. From FIG. 11, it can be seen that the performance can be greatly improved by changing SM to HSM. In addition, by performing competitive learning between classes using mirror image learning, performance superior to conventional DNN / HMM; bi-gram is obtained even without a language model. Furthermore, it can be seen that the performance can be improved by adding a language model to bi-gram and tri-gram.

  As mentioned above, although embodiment of this invention was described, the said embodiment was only shown as an example to the last and is not intending limiting the range of invention. The embodiments presented herein can be implemented in various other forms, and various omissions, substitutions, and changes can be made without departing from the spirit of the invention.

  The pattern classification apparatus, pattern classification method, and pattern classification program of the present invention can accurately extract, for example, phonemes and articulatory features from many patterns (vector data of speech, characters, faces, etc.). The amount of calculation for processing the data constituting the can be significantly reduced. For this reason, the pattern classification apparatus, pattern classification method, and pattern classification program of the present invention can contribute to the development of the information processing industry in which it is essential to process many patterns (vector data such as speech, characters, and faces). it can. The pattern classification apparatus, pattern classification method, and pattern classification program of the present invention recognize, classify, and discriminate many patterns (vector data such as speech, characters, and faces) with high accuracy. Therefore, it can be considered that this will greatly contribute to the development of the robot-related technology industry.

DESCRIPTION OF SYMBOLS 1 Classifier 2 Preprocessing part 3 Eigenvector set 4 Orthogonalization vector reconstruction part 5 Multiple linear mapping part 6 Class-specific eigenvector set 7 Higher order subspace similarity calculation part 8 Similarity posterior probability processing part 10 Feature classification part (HMM phoneme Classification part)

Claims (9)

A pattern classification device capable of recognizing an input pattern by extracting a pattern feature similarity vector,
A preprocessing unit for performing preprocessing on the input pattern;
A high-dimensional vector conversion unit that converts the pre-processed vector into a high-dimensional vector by multiple linear mapping (tensor product);
A high-order subspace similarity calculation unit that generates and outputs a high-dimensional eigenvector set from the high-dimensional vector, and generates a class-wise similarity between the high-dimensional vector and the high-dimensional eigenvector set Ψ; ,
A pattern classification apparatus comprising: a feature classification unit that classifies individual features based on the similarity by class.   The pre-processing unit maps the input pattern to a class-common dual space by taking an inner product of the input pattern and a common eigenvector set common to the feature class previously generated by principal component analysis on the learning pattern, and the inner product 2. The pattern classification apparatus according to claim 1, further comprising a process of reconstructing an orthogonal vector common to the classes by arranging the scalar values obtained from the operations and arranging the number of eigenvector axes. The higher-order subspace similarity calculation unit has a competitive learning unit that corrects classification error by applying competitive learning using a mirror image to the subspace method,
Class-specific eigenvectors Ψ (k, m) obtained after competitive learning including mirror image learning between (auto) correlation matrices for each class using high-dimensional feature learning data in the competitive learning unit, The pattern classification apparatus according to claim 1, wherein the pattern classification apparatus is used as a higher-order subspace similarity calculation as an eigenvector set.   The classification device according to claim 1, further comprising a similarity posterior probability processing unit, wherein the similarity posterior probability processing unit receives the similarity as an input and calculates a prior probability of a class to which the input vector data belongs. A pattern classification device characterized by having a configuration to perform. A pattern classification method capable of recognizing an input pattern by extracting a pattern feature similarity vector,
A preprocessing step of performing preprocessing on the input pattern;
A high-dimensional vector conversion step of converting the pre-processed vector into a high-dimensional vector by multiple linear mapping (tensor product);
A high-order subspace similarity calculation step for generating and outputting a high-dimensional eigenvector set from the high-dimensional vector, and calculating and generating a class-wise similarity between the high-dimensional vector and the high-dimensional eigenvector set Ψ; ,
A pattern classification method comprising: a feature classification step of classifying individual features from the similarity by class.   The preprocessing step maps the input pattern to a class-common dual space by taking an inner product of the input pattern and an eigenvector set common to a feature class previously generated by principal component analysis on the learning pattern, and the inner product The pattern classification method according to claim 5, further comprising a process of reconstructing an orthogonal vector common to the classes by arranging the scalar values obtained from the operations and arranging the number of eigenvector axes. The higher-order subspace similarity calculation step includes a competitive learning step of correcting classification error by applying competitive learning using a mirror image to a subspace method,
In the competitive learning step, class-specific eigenvectors Ψ (k, m) obtained after performing competitive learning including mirror image learning between (auto) correlation matrices for each class using high-dimensional feature learning data, 6. The pattern classification method according to claim 5, wherein the eigenvector set is used for high-order subspace similarity calculation.   6. The classification method according to claim 5, further comprising a similarity posterior probability processing step, wherein the similarity posterior probability processing step receives the similarity as an input and calculates a prior probability of a class to which the input vector data belongs. A pattern classification method characterized by: The pattern classification program for making a computer perform the pattern classification method as described in any one of Claims 5-8.