JP7019138B2 - Coding device, coding method and program - Google Patents

Description

The present invention relates to a coding apparatus and a coding method, and a program for executing the coding method.

In recent years, methods using deep learning have dramatically improved accuracy, and have been actively researched and used in a wide range of fields such as image recognition and voice recognition. Although many deep learning methods have been proposed so far, the restricted Boltzmann machine (hereinafter referred to as "RBM") is used as the most representative model. In addition, Deep Belief Net (hereinafter referred to as "DBN") in which RBMs are stacked in multiple layers is also used. In addition, various RBM extensions have been proposed.

“Lending Direction to Neural Networks”: Neural Networks Vol.8. No4.pp503-512,1995 (Richard S. Zemel, Christopher K. Williams, Michael C. Mozer)

Conventionally, as a feature amount extraction process using RBM, a binary or a real value has been used as an input feature amount in either approach.
For example, when performing speech processing such as speech recognition and speech synthesis, acoustic features based on amplitude spectra such as Mel-Frequency Cepstrum Coefficients (MFCC), Mel-Frequency Cepstrum Coefficients, and STRAIGHT spectra are used. ing. However, in the acoustic feature amount extraction based on the amplitude spectrum, the phase information is missing, and there is not a little loss of information with respect to the original complex-numbered voice data.
Here, speech processing has been described as an example, but there is a problem that information loss exists even when feature quantity extraction is performed from other complex number information.

Note that Non-Patent Document 1 describes a technique for extracting a feature amount using a complex number with a Boltzmann machine, but since this technique does not apply the above-mentioned RBM or DBN, the feature amount is extracted. It has been desired to develop a method that can be performed more accurately.

The present invention provides a coding device, a coding method, and a program capable of performing good coding based on the feature extraction by applying RBM to a complex number and performing feature extraction with high accuracy. The purpose.

The coding device of the present invention includes a parameter learning unit and a coding unit.
The parameter learning unit applies a probabilistic model based on a restricted Boltzmann machine that assumes that there is a coupling weight between the visible element that represents the input data and the hidden element that represents the potential information, and applies it to the training data. Then, the process of estimating the hidden element and the coupling weight is performed.
The coding unit applies a probability model by the restricted Boltzmann machine estimated by the parameter learning unit to the coding input data, estimates hidden elements, and outputs the estimated hidden elements as coding data.
Here, the training data and the coding input data are complex number data, and are characterized in that the energy function of the probability model by the restricted Boltzmann machine includes a cross-term of a real part and an imaginary part .

Further, the coding method of the present invention includes a parameter learning process and a coding process.
The parameter learning process applies a probabilistic model based on a restricted Boltzmann machine that assumes that there is a coupling weight between the visible element that represents the input data and the hidden element that represents the potential information, and applies the probability model to the training data. Then, the process of estimating the hidden element and the coupling weight is performed.
In the coding process, a stochastic model by the restricted Boltzmann machine estimated by the parameter learning process is applied to the coding input data to estimate the hidden element, and the estimated hidden element is output as the coded data.
Here, the training data obtained by the parameter learning process and the coding input data obtained by the coding process are complex number data, and the energy function of the probability model by the restricted Boltzmann machine includes the cross-term of the real part and the imaginary part. It is characterized by being.

Further, in the program of the present invention, a computer performs a step of executing the parameter learning process of the above-mentioned coding method and a step of executing the coding process by inputting the learning data and the coding input data composed of complex number data. Is to be executed.

According to the present invention, it is possible to extract the feature amount by the complex RBM which is an extension of the restricted Boltzmann machine (RBM) to a complex number, and it is possible to extract and encode the feature amount from the input data with high accuracy. , Efficient coding will be possible.

It is a block diagram which shows the structural example of the coding apparatus by one Embodiment of this invention. It is a block diagram which shows the hardware configuration example of the coding apparatus of FIG. It is a figure which shows typically the complex RBM (Restricted Boltzmann machine) which is the probability model applied to one Embodiment of this invention. It is a flowchart which shows the flow of parameter learning by one Embodiment of this invention. It is a flowchart which shows the flow of coding by one Embodiment of this invention. It is a flowchart which shows the learning process of the complex RBM of step S13 of FIG. It is a flowchart which shows the coding process of step S23 of FIG. It is a block diagram which shows the structural example of the decoding apparatus which decodes the data encoded by one Embodiment of this invention. It is a flowchart which shows the flow of decoding by one Embodiment of this invention. It is a flowchart which shows the decoding process of step S52 of FIG. It is a figure which shows the example of the original data (FIG. 11A), and the coded data (FIG. 11B) to which one Embodiment of this invention is applied. It is a characteristic diagram which compared the reconstruction error by the complex RBM to which this invention was applied, and the reconstruction error by the conventional example (GB-RBM). It is a figure which shows typically the example in which the complex RBM applied to one Embodiment of this invention is multi-layered.

Hereinafter, an example of a preferred embodiment of the present invention will be described.

[1. Configuration example of coding device]
FIG. 1 is a diagram showing a configuration example of a coding device according to an embodiment of the present invention. As shown in FIG. 1, the coding device 1 configured by a computer (PC) or the like includes a parameter learning unit 11 and a coding processing unit 12.
The parameter learning unit 11 performs learning processing in advance on the same type of data as the data to be encoded, and obtains the parameters required for encoding. The coding processing unit 12 encodes the input data (coding data) by using the parameters obtained in the learning processing.
As the input data to be encoded, various data such as voice data and image data can be applied. However, as will be described later, the learning data and the input data handled in this embodiment are complex number data.

The parameter learning unit 11 includes a complex number data acquisition unit 111, a preprocessing unit 112, and a parameter estimation unit 113. The complex number data for learning is supplied to the complex number data acquisition unit 111. The learning complex number data acquired by the complex number data acquisition unit 111 is supplied to the parameter estimation unit 113 after being preprocessed by the preprocessing unit 112.
For example, when the learning complex number data acquired by the complex number data acquisition unit 111 is voice data, the preprocessing unit 112 cuts out the training voice data every unit time (hereinafter referred to as a frame) and MFCC (hereinafter referred to as a frame). Calculates and normalizes the spectral features of the audio signal for each frame, such as Mel-Frequency Cepstrum Coefficients) and Mel-Frequency Cepstrum features. The training data may be converted into complex number data by the processing in the preprocessing unit 112.

The parameter estimation unit 113 has a probability model composed of a visible element estimation unit 1131 and a hidden element estimation unit 1132. In the present embodiment, a complex RBM (Complex RBM) obtained by extending RBM to a complex number is used as a probability model composed of the visible element estimation unit 1131 and the hidden element estimation unit 1132. In addition to the visible element and the hidden element, the probability model of the complex RBM also has information on the coupling weight between the elements, and the parameter estimation unit 113 also estimates and possesses the information on the coupling weight. The details of this complex RBM will be described later.

The coding processing unit 12 includes a complex number data acquisition unit 121, a preprocessing unit 122, and a coding unit 123.
The complex number data acquisition unit 121 is supplied with the complex number data for coding. The complex number data for coding acquired by the complex number data acquisition unit 121 is supplied to the coding unit 123 after being preprocessed by the preprocessing unit 122.
The preprocessing unit 122 has the same configuration as the preprocessing unit 112 of the parameter learning unit 11. The coding data may be converted into complex number data by the processing in the preprocessing unit 122.

The coding unit 123 has the same configuration as the parameter estimation unit 113 of the parameter learning unit 11, and is a complex RBM composed of a visible element obtained by the visible element estimation unit 1231 and a hidden element obtained by the hidden element estimation unit 1232. It has a probabilistic model. When the visible element estimation unit 1231 and the hidden element estimation unit 1232 estimate the visible element and the hidden element, the parameters estimated by the parameter estimation unit 113 of the parameter learning unit 11 are used.

The coding device 1 outputs the hidden element estimated by the hidden element estimation unit 1232 of the coding unit 123 to the outside as coded data.
In the configuration shown in FIG. 1, the parameter estimation unit 113 that performs learning processing and the coding unit 123 that performs input data coding processing are individually configured, but the parameter estimation unit 113 and the coding unit 123 have different configurations. It has almost the same function, and the parameter estimation unit 113 may process the coding unit 123. The complex number data acquisition units 111 and 121 and the preprocessing units 112 and 122 may also be shared.

FIG. 2 is a diagram showing a hardware configuration example of the coding device 1. Here, an example in which the coding device 1 is configured by a computer (PC) is shown.
As shown in FIG. 2, the coding apparatus 1 has a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, and a RAM (Random Access Memory) 103, which are connected to each other via a bus 107. , HDD (Hard Disk Drive) / SSD (Solid State Drive) 104, connection I / F (Interface) 105, communication I / F 106. The CPU 101 comprehensively controls the operation of the coding device 1 by executing a program stored in the ROM 102, the HDD / SSD 104, or the like with the RAM 103 as a work area. The connection I / F 105 is an interface with a device connected to the coding device 1. The communication I / F is an interface for communicating with other information processing devices via a network.

The input / output and setting of the learning data and the coding data are performed via the connection I / F 105 or the communication I / F 106. The function of the coding apparatus 1 described with reference to FIG. 1 is realized by executing a predetermined program in the CPU 101. The program may be acquired via a recording medium, may be acquired via a network, or may be incorporated into a ROM for use. In addition, hardware for realizing the configuration of the encoding device 1 by building a logic circuit such as an ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array) instead of a general computer and program combination. It may be configured.

[2. Definition of complex RBM]
Next, the complex RBM, which is a probabilistic model of the parameter estimation unit 113 and the coding unit 123, will be described.
The RBM has assumed that there is a bidirectional connection weight between the visible element representing the input data and the hidden element representing the potential information (although there is no connection between the visible or hidden elements). It is a stochastic model, and the complex RBM is an extension of the RBM to a complex number having a real part and an imaginary part.
FIG. 3 shows a graph representation example of the complex RBM of the present embodiment.
The example of FIG. 3 shows a model of a complex RBM whose visible element is ^I -dimensional data z ∈ CI which is a complex number.
In FIG. 3, z is a visible element, h is a hidden element, W'is a bidirectional coupling weight between the visible element z and the hidden element h, b'is the bias of the visible element z, and c is the hidden element h. Bias, q indicates conjugation. The line (overline) attached above each sign indicates the complex conjugate.

This complex RBM is defined by the following equations [Equation 1] to [Equation 4]. Here, the ^I -dimensional data z ∈ CI is a visible element, the set of parameters of the probability model is θ, and the superscript H indicates the Hermitian transpose.

Further, Φ in the equation [Equation 3] is defined by the equation [Equation 5], and the parameter representing the variance and pseudo-variance (covariance with the conjugate complex number) of the complex number Z defined in the equation [Equation 5] is [. It is defined by the equation [Equation 6]. However, Δ is a function that returns a diagonal matrix in which the input vector is a diagonal component.

After all, the parameter of the complex RBM is θ = {b, c, W, γ, δ}. Here, the [Equation 7] equation and the [Equation 8] equation are introduced. However, in the [Equation 7] and [Equation 8] equations, the fractional line represents element division.

From this, it becomes the formula [Equation 9].

The energy function defined by the equation [Equation 3] can be rewritten into the equation [Equation 10]. R is a function that returns the real part of the input complex number.

Here, the energy function is a real value. It can be confirmed that each dimension of the complex visible element z has a coupling with a conjugate complex number, but there is no coupling between dimensions like a normal RBM (non-complex RBM). Further, by using the following [Equation 11] and [Equation 12] equations, the [Equation 3] equation becomes the [Equation 13] equation.

From this equation [Equation 13], as shown in FIG. 3, it can be seen that the relationship between z and h and the relationship between z ( ⁻ ) and h are mirror images of each other with the conjugate space in between. The "( ^- )" of "z ( ^- )" shown in the present specification is an overline indicating complex conjugate, and originally, "-" is above "z" as shown in FIG. Although it is added, in this specification, it is described as "z ( ^- )" due to the limitation of description. Overlines added to other symbols are described herein as well.
From the above definitions, the conditional probability of the visible element when the hidden element is given and the conditional probability of the hidden element when the visible element is given are given by the equations [Equation 14] and [Equation 15], respectively. Can be represented.

However, CN (・; μ, Γ, C) is defined by the [Equation 16] and [Equation 17] equations, which are multivariate complex normal distributions of the mean μ, the variance-covariance matrix Γ, and the pseudovariance-covariance matrix C. To. B (・; π) represents a multidimensional Bernoulli distribution with a success probability of π. f (・) represents a sigmoid function for each element. D is the number of dimensions of z.

[3. Learning processing operation and coding processing operation]
Next, the coding process performed by applying the complex RBM of the present embodiment will be described.
FIG. 4 is a flowchart showing the flow of the parameter learning operation performed by the parameter learning unit 11.
First, the complex number data acquisition unit 111 acquires the complex number data for learning (step S11), and the preprocessing unit 112 executes the preprocessing of the complex number data (step S12). For example, when the learning data is voice data, the complex number data acquisition unit 111 cuts out the learning voice data for each frame (for example, every 5 msec), and performs FFT processing on the cut out learning voice signal. By applying, the spectral feature amount (for example, MFCC or mercepstrum feature amount) is calculated. In this preprocessing, the training data may be used as complex number data.

Next, the preprocessed complex number data is supplied to the parameter estimation unit 113, and the parameter estimation unit 113 performs the parameter learning process of the complex number data (step S13). The details of the parameter learning process performed in step S13 will be described later (FIG. 6).
In this parameter learning process, each parameter of the complex RBM model is determined and stored. Then, the stored parameter is passed to the coding unit 123 and encoded by the coding unit 123 (step S14).

FIG. 5 is a flowchart showing the flow of the coding process performed by the coding process unit 12.
First, the complex number data acquisition unit 121 acquires the complex number data for coding (step S21), and the preprocessing unit 122 executes the preprocessing of the complex number data (step S22). The pre-processing here is the same as the pre-processing in step S12 performed by the pre-processing unit 112. As described earlier in the configuration of the preprocessing unit 112, the input data may be converted into complex number data by this preprocessing.

The preprocessed complex number data is supplied to the coding unit 123, and the coding unit 123 estimates the hidden element using the parameters of the complex RBM model passed in step S14 and performs the coding process. (Step S23). Details of the coding process performed in step S23 will be described later (FIG. 7). Then, the coding processing unit 12 outputs the hidden element obtained in step S23 as coding data (step S24).

FIG. 6 is a flowchart showing the details of the parameter learning process performed in step S13 of FIG.
First, the parameter estimation unit 113 sets an arbitrary value as a parameter of the complex RBM model (step S31). Next, the preprocessed complex number data for learning is input to the visible element estimation unit 1131 of the parameter estimation unit 113 (step S32).
After that, the parameter estimation unit 113 calculates the probability value of the hidden element of the complex RBM model and samples the calculated value (step S33). Here, "sampling" means randomly generating one piece of data according to the conditional probability density function, and is used hereinafter with the same meaning.

Further, the parameter estimation unit 113 calculates the probability value of the visible element of the complex RBM model, samples the calculated value (step S34), and then recalculates and calculates the probability value of the hidden element of the complex RBM model. The value is resampled (step S35). Then, the parameter estimation unit 113 updates the various parameters obtained by the calculations so far as parameters constituting the model of the complex RBM, and stores the updated values (step S36).

After updating the parameters in step S36, the parameter estimation unit 113 determines whether or not the end condition of the parameter learning process is satisfied (step S37), and if it is determined that the end condition is not satisfied (NO in step S37). ), Return to step S31, and repeat the process up to this point. If it is determined in step S37 that the end condition is satisfied (YES in step S37), the parameter estimation unit 113 ends the parameter learning process. The end condition in step S37 includes, for example, the number of repetitions of these series of steps.

FIG. 7 is a flowchart showing details of the coding process performed in step S33 of FIG.
First, the coding unit 123 sets the parameters passed from the parameter estimation unit 113 (step S41). Next, the preprocessed complex number data for coding is input to the visible element estimation unit 1231 of the coding unit 123 (step S42).
After that, the hidden element estimation unit 1232 of the coding unit 123 calculates the hidden element of the complex RBM model and outputs the estimated hidden element as coding data (step S43).

Next, the parameter estimation process of the complex RBM model, which is performed in the specific learning process and the coding process, will be described using mathematical formulas.
In the parameter estimation, the parameters of the complex RBM are updated by the complex gradient method so as to maximize the log-likelihood L (θ) of the input data (visible data) z represented by the following equation [Equation 18]. Variables with tildes were introduced to distinguish them from variables without tildes.

In the complex gradient method, the parameters are updated by repeatedly executing the calculation of the equation [Equation 19] using the learning rate α> 0.

However, the partial derivative of the complex number in the equation [Equation 19] is the derivative of Weltinger shown in the equation [Equation 20]. Here, i is an imaginary unit. [Equation 20] The first and second terms on the right side of the equation represent the partial differential of the log-likelihood L with respect to the real part and the partial differential with respect to the imaginary part of the parameter θ, respectively.

The partial derivative of each parameter includes terms for the expected value for the observed data (input data) and the expected value of the model. Since the expected value of the model is difficult to calculate, it is approximately calculated using the CD method (Contrastive Divergence method) in the same manner as in the case of calculating with the conventional RBM.
The partial differential of the parameter with respect to the energy function can be obtained analytically, and is shown in the equations [Equation 21] to [Equation 25], respectively.

However, ◯, |, |, and ² represent the product, the absolute value, and the square of each element, respectively, and are represented by the following equations [Equation 26] and [Equation 27].

Since the variance and pseudo-variance updates have different scales compared to other parameters, in order to train them stably, they are actually replaced as shown in [Equation 28], and the parameters are updated with r and s. ..

[4. Decryptor configuration and operation]
FIG. 8 shows a configuration example of the decoding device 2 corresponding to the coding device 1 according to the embodiment of the present invention.
The decoding device 2 decodes the coded data obtained by the coding device 1, and is configured by, for example, a computer. The decoding device 2 may be integrated with the coding device 1.
The decoding device 2 includes a parameter learning unit 11 and a decoding processing unit 13.
The parameter learning unit 11 is the same as the parameter learning unit 11 of the coding device 1, and as the parameter estimation unit 113, the visible element estimation unit 1131 and the hidden element estimation unit 1132 for estimating the visible element and the hidden element obtained in the learning process are used. Be prepared.

The coding data obtained by the coding device 1 is supplied to the decoding processing unit 13. The decoding processing unit 13 includes a decoding unit 131. The decoding unit 131 has a visible element estimation unit 1311 and a hidden element estimation unit 1312, and acquires the parameters of the complex RBM model from the parameter estimation unit 113.
The hidden element estimation unit 1312 uses the input coded data as a hidden element. Then, the visible element estimation unit 1311 obtains an estimated value of the visible element by an operation using the parameters of the complex RBM model. The estimated value of this visible element is supplied to the post-processing unit 132, and the post-processing unit 132 performs post-processing. In the post-processing unit 132, for example, a process of restoring the pre-processing in the pre-processing unit 122 of the coding device 1 is performed.
Then, the output unit 133 outputs the decrypted data after the post-processing.

FIG. 9 is a flowchart showing the flow of decoding by the decoding device 2.
When the decoding device 2 acquires the coded data to be decoded (step S51), the decoding processing unit 13 performs the decoding process. Details of the decryption process will be described later (FIG. 10).
The data obtained by the processing in the decoding processing unit 13 is supplied to the post-processing unit 132 and post-processed (step S52), and the post-processed data is output as decryption data from the output unit 133. (Step S53).

FIG. 10 shows the details of the decoding process in step S52 of the flowchart of FIG.
First, the decoding unit 131 sets various parameters of the complex RBM model passed from the parameter learning unit 11 (step S61). Here, the parameters used when coding the coded data to be decoded (parameters used at the time of coding in the coding device 1 shown in FIG. 1) are acquired from the parameter learning unit 11 and set. Then, the coded data is input to the hidden element estimation unit 1312 of the decoding unit 131 (step S62). Then, the visible element estimation unit 1311 estimates the visible element (decoding data) using the model of the complex RBM (step S63).
In this way, the coded data can be decoded in the reverse flow of the coding.

[5. Differences between complex RBM and conventional method (GB-RBM)]
The complex number z = x + iy is GB-RBM (Gaussian-Bernoulli RBM), which is one of the conventional methods, by using z'= [x ^T y ^T ] ^T ∈ R ^2I , which is a connecting vector of the real part and the imaginary part. It can also be expressed by. GB-RBM is represented by the following equations [Equation 29] to [Equation 32].

However, Σ _x = Δ (σ _x ² ) and Σ _y = Δ (σ _y ² ). In this case, for example, the partial differentials of the bias parameters of the real part and the imaginary part with respect to the energy function are shown by the equations [Equation 33] and [Equation 34], respectively.

On the other hand, if z = x + iy, b = b ^R + ib ^I , W = ^WR + ^iWI , q = q ^R + iq ^I , and the energy function of the complex RBM (the right side of the equation [Equation 3]) is rewritten, [Equation 35] ] Expression.

However, the conditions shown in the following equations [Equation 36] to [Equation 42] were set.

Here, when the equation [Equation 31] and the equation [Equation 35] are compared, the cross-term of x and y (x ^T Σ _xy ^-1 y) is included in the modeling by the complex RBM according to the embodiment of the present embodiment. You can see that there is. That is, it can be said that the complex RBM is an extended expression that considers the relationship between the real part and the imaginary part for each feature dimension, in addition to the complex expression by GB-RBM, which is one of the conventional methods.
Furthermore, in the complex representation by GB-RBM, the biases of the real part and the imaginary part of the observed data are calculated independently (for example, updating the real part bias) as shown by the equations [Equation 33] and [Equation 34]. In (the information of only the real part is used), in the update formula of the bias parameter of the complex RBM ([Equation 21] formula), both the real part and the imaginary part are used and updated. Therefore, in the modeling by the complex RBM according to the embodiment of the present embodiment, the learning can be performed while maintaining the data structure of the complex number.

[6. Experimental example]
Next, an example experimented to verify the validity of the model by the complex RBM according to the present embodiment will be described.
Here, in order to confirm the validity of the model by the complex RBM according to the embodiment of the present embodiment, voice data was coded and a quality evaluation experiment of the coded voice was performed. Specifically, we conducted a quality evaluation experiment of reconstructed speech using the Repeated Harvard Sentence Prompts (REHASP) ² corpus, and used one repeat of speech (30 sentences, about 20 seconds, sampling rate 16 kHz) from the corpus. .. Then, a complex RBM having 200 hidden elements was trained using a complex spectrum (129 dimensions) subjected to a short-time Fourier transform with a window width of 256,64 sample overlap as a visible element. At this time, a stochastic gradient descent method with a learning rate of 0.01, a moment coefficient of 0.1, a batch size of 100, and a number of repetitions of 100 was used. Further, as a comparison method, GB-RBM (the number of hidden elements is 200) in which a vector connecting the real part and the imaginary part of the same complex spectrum data is used as a visible element was trained under the same conditions.

FIG. 11 is a diagram comparing the original amplitude spectrum before coding (FIG. 11A) with the spectrum restored by the model by the complex RBM according to the embodiment of the present embodiment (FIG. 11B). In FIG. 11, the vertical axis represents frequency and the horizontal axis represents time. As can be seen from FIG. 11, the spectrum restored by the model by the complex RBM according to the present embodiment is close to the original spectrum, and the complex RBM according to the present embodiment encodes and decodes the voice spectrum with high accuracy. Can be confirmed to be possible.

FIG. 12 shows a comparison between the complex RBM (Comp RBM) according to the present embodiment and the conventional RBM (GB-RBM) based on the reconstruction error during learning, and the vertical axis indicates the number of reconstruction errors. , The horizontal axis shows the change over time. "Adam" or "Ada Grad" is the result when Adam or Ada Grad is used as the optimization method, respectively, and the one without notation shows the result when the stochastic gradient descent method is used as the optimization method.
In FIG. 12, an example in which a complex RBM (Comp RBM) is applied alone, an example in which it is combined with another method (Comp RBM + Ada Grad, Comp RBM + Adam), an example in which a conventional RBM is applied alone, and another method are shown. Six examples with combined examples (RBM + Ada Grad, RBM + Adam) are shown.
For example, the characteristics of the example of combining Adam with complex RBM [Comp RBM + Adam: thick solid line] converge faster than the characteristics of the example of combining Adam with conventional RBM [RBM + Adam: thin solid line], and the error at the time of convergence is low. You can see that. The characteristics of the example [Comp RBM: two-dot chain line] to which only the complex RBM is applied also converge faster than the characteristics of the example [RBM: one-dot chain line] to which only the conventional RBM is applied, and the error at the time of convergence is low. I understand.

[7. Modification example]
Although the model of the complex RBM shown in FIG. 3 shows the configuration of a single layer RBM, the complex RBM of the present invention may be applied to a DBN (Deep Belief Net) in which RBMs are stacked in multiple layers.
FIG. 13 shows an example in which the complex RBM is three-layered.
The real part obtains the hidden code h ₁ of the first layer from the visible element z and the bias c ₁ of the hidden element h. The imaginary portion obtains the hidden code h ₁ of the first layer from the visible element z ( ⁻ ) and the bias c ₁ of the hidden element h. W ₁ ′ and W ₁ ′ ( ⁻ ) are bidirectional coupling weights between the visible element z and the hidden element h ₁ .
From the hidden sign h ₁ and the bias c ₁ of the real part of the first layer, the hidden sign h ₂ and the bias c ₂ of the second layer are obtained, and from the hidden sign h ₁ and the bias c ₁ of the real part of the first layer, 2 The hidden sign h ₂ and the bias c ₂ of the layer are obtained. W ₂ ′ and W ₂ ′ ( ⁻ ) are bidirectional coupling weights between the hidden element h ₁ and the hidden element h ₂ .
Further, from the hidden code h ₂ and the bias c ₂ of the real part of the second layer, the hidden code h ₃ and the bias c ₃ of the third layer are obtained, and from the hidden code h ₂ and the bias c ₂ of the real part of the second layer. The hidden sign h ₃ and the bias c ₃ of the third layer are obtained. W ₃ ′ and W ₃ ′ ( ⁻ ) are bidirectional coupling weights between the hidden element h ₂ and the hidden element h ₃ .
In this way, even with the multi-layered complex RBM, coding and decoding can be performed in the same manner.

Further, in the above-described embodiment, the case of applying to voice data has been described as an experimental example, but the complex RBM according to the present invention can be applied to coding and decoding of various other signals. For example, the complex RBM according to the present invention may be applied to the coding and decoding of image data. Further, the complex RBM according to the present invention may be applied to coding and decoding of data other than audio data and image data.

1 ... Coding device, 2 ... Decoding device, 11 ... Parameter learning unit, 12 ... Coding processing unit, 13 ... Decoding processing unit, 101 ... CPU (central control) Unit), 102 ... ROM, 103 ... RAM, 104 ... HDD / SDD, 105 ... Connection I / F, 106 ... Communication I / F, 111, 121 ... Complex number data acquisition Units 112, 122 ... Pre-processing unit, 113 ... Parameter estimation unit, 123 ... Encoding unit, 131 ... Decoding unit, 132 ... Post-processing unit, 133 ... Output unit , 1131, 1231, 1311 ... Visible element estimation unit, 1132, 1232, 1312 ... Hidden element estimation unit

Claims (5)

The hidden element is applied to the learning data by applying a probabilistic model by a restricted Boltzmann machine assuming that there is a coupling weight between the visible element representing the input data and the hidden element expressing the potential information. And a parameter learning unit that performs processing to estimate the connection weight,
A coding unit that estimates the hidden element by applying the probability model by the restricted Boltzmann machine estimated by the parameter learning unit to the input data for coding, and outputs the estimated hidden element as coding data. And with
The training data and the coding input data are complex number data, and the energy function of the probability model by the restricted Boltzmann machine includes a cross-term of a real part and an imaginary part.
Coding device. The probability model by the restricted Boltsman machine has a visible element z composed of I-dimensional data z ∈ CI and a hidden element h, the parameter set of the model is θ, and the parameters constituting the parameter set θ are b, c, and so on. W, γ, and δ, the bias of the visible element is b ∈ CI, the bias of the hidden element is c ∈ RJ, the complex conjugate weight between the visible element and the hidden element is W ∈ CI × J, and the overline of each sign is Complex conjugate, H is defined by the following equation when it is Hermitian translocation.

The coding device according to claim 1. The coding device according to any one of claims 1 to 2, further comprising a decoding processing unit for decoding the coded data obtained by the coding unit. The hidden element is applied to the learning data by applying a probabilistic model by a restricted Boltzmann machine assuming that there is a coupling weight between the visible element representing the input data and the hidden element expressing the potential information. And the parameter learning process in which the arithmetic processing unit executes the process of estimating the join weight, and
The arithmetic processing unit executes the process of estimating the hidden element by applying the probability model by the restricted Boltzmann machine estimated by the parameter learning process to the input data for coding, and the estimated hidden element is coded. Coding processing to output as conversion data and
Including
The learning data obtained by the parameter learning process and the coding input data obtained by the coding process are complex number data, and the cross-term of the real part and the imaginary part in the energy function of the probability model by the restricted Boltzmann machine. It is included
Coding method. The hidden element is applied to the input data by applying a probabilistic model by a restricted Boltzmann machine assuming that there is a coupling weight between the visible element representing the input data and the hidden element expressing the potential information. And the parameter learning step of obtaining the training data which is the complex number data by performing the process of estimating the connection weight.
A code that estimates the hidden element by applying a probability model by the restricted Boltzmann machine estimated in the parameter learning step to the input data, and outputs the estimated hidden element as coded data which is complex data. Including the conversion step,
The energy function of the stochastic model by the restricted Boltzmann machine contains the cross-term of the real part and the imaginary part.
A program that causes a computer to execute each of the above steps.

Images