US 7,203,368 B2Grant

Embedded bayesian network for pattern recognition

Issue Date:2007-04-10

•18 Claims

•5 Drawing Sheets

Abstract

A pattern recognition procedure forms a hierarchical statistical model using a hidden Markov model and a coupled hidden Markov model. The hierarchical statistical model supports a pa 20 layer having multiple supernodes and a child layer having multiple nodes associated with each supernode of the parent layer. After training, the hierarchical statistical model uses observation vectors extracted from a data set to find a substantially optimal state sequence segmentation.

Metadata

Assignee

Intel Corporation

Inventor

Ara V. Nefian

Application Information

Application Number:US 10/269,381

Filing Date:2003-01-06

Priority Date:2003-01-06

Art Unit:2624

Classifications

IPC:

G06K9/62G06K9/00G10L15/00G05B13/02

Patent Drawings (5 sheets)

Description

Field of the Invention

[0001] The present invention relates to computer mediated pattern detection. More particularly, the present invention relates to improved Bayesian networks for classifying data.

Background

[0002] Bayesian networks such as those represented by the hidden Markov model (HMM) and coupled Hidden Markov (CHMM) models have long been used to model data for the purposes of pattern recognition. Any discrete time and space dynamical system governed by such a Bayesian network emits a sequence of observable outputs with one output (observation) for each state in a trajectory of such states. From the observable sequence of outputs, the most likely dynamical system can be calculated. The result is a model for the underlying process. Alternatively, given a sequence of outputs, the most likely sequence of states can be determined.

[0003] For example, one dimensional HMMs have been widely used in speech recognition to model phonemes, words, or even phrases, and two dimensional HMMs have been used for image processing tasks. One of the important characteristics of a HMM is its ability to cope with variations in feature space, allowing data modeling with variations along different dimensions. Coupled hidden Markov models can be similarly employed, since they correspond to a generalization of a HMM. A CHMM may comprise a collection of HMMs, each of which corresponds to a data channel.

Brief Description of the Drawings

[0004] The inventions will be understood more fully from the detailed description given below and from the accompanying drawings of embodiments of the inventions which, however, should not be taken to limit the inventions to the specific embodiments described, but are for explanation and understanding only.

[0005] FIG. 1 schematically illustrates a data classification system;

[0006] FIG. 2 generically illustrates an embedded hidden Markov model-coupled hidden Markov model (HMM-CHMM) structure;

[0007] FIG. 3 is a flow diagram illustrating training of an embedded HMM-CHMM; and

[0008] FIG. 4 generically illustrates an embedded coupled hidden Markov model-hidden Markov model (CHMM-HMM) structure.

Detailed Description

[0009] FIG. 1 generally illustrates a system 10 for data analysis of a data set 12 using an embedded Bayesian network that includes a hidden Markov model (HMM) and a coupled hidden Markov model (CHMM). A embedded Bayesian network is used because it has good generalization performance even for high dimensional input data and small training sets.

[0010] The data set 12 can include static or video imagery 14 containing objects to be identified or classified, including but not limited to textual characters, ideographs, symbols, fingerprints, or even facial imagery 15. In addition, non-image data sets such as bioinformatic databases 16 containing, for example, gene or protein sequences, DNA microarray data, sequence data, phylogenetic information, promoter region information; or textual, linguistic, or speech analysis data suitable for machine learning/identification 18 can be used. The same data set can be optionally used both to train and classify data with the appropriate training module 20 and classification module 22.

[0011] The processing procedure for system 10 maybe performed by a properly programmed general-purpose computer alone or in connection with a special purpose computer. Such processing may be performed by a single platform or by a distributed processing platform. In addition, such processing and functionality can be implemented in the form of special purpose hardware, custom application specific integrated circuits (ASICs), configurable FPGA circuits, or in the form of software or firmware being run by a general-purpose or network processor. Data handled in such processing or created as a result of such processing can be stored in any memory as is conventional in the art. By way of example, such data may be stored in a temporary memory, such as in the RAM of a given computer system or subsystem. In addition, or in the alternative, such data may be stored in longer-term storage devices, for example, magnetic disks, rewritable optical disks, and so on. For purposes of the disclosure herein, a computer-readable media may comprise any form of data storage mechanism, including such existing memory technologies as well as hardware or circuit representations of such structures and of such data.

[0012] FIG. 2 generically illustrates a logical structure 30 of an embedded hidden Markov model-coupled hidden Markov model (HMM-CHMM). As seen in FIG. 2, HMM-CHMM is a hierarchical statistical model that includes a HMM parent layer 32 (collectively formed from nodes 33) and a CHMM child layer 34 (collectively formed from nodes 35). The child layer 34 associates one CHMM node 35 to each node 33 in the parent layer 32, and the parameters of the individual CHMMs remain independent from each other. Instead, the parameters of each child layer CHMM depend upon the state of the connected parent node 33. Typically, for multidimensional data sets, the HMM in the parent layer 32 is associated with at least one dimension, and the CHMM child layers are associated with data in an orthogonal dimension with respect to the parent layer.

[0013] Formally defined, the elements of an embedded HMM-CHMM have:an initial super state probability π_0,0and a super state transition probability from super state j to super state i, a_0,t|j, where super state refers to the state of the parent layer 32 HMM node 33.

[0014] For each super state k the parameters of the corresponding CHMM are defined to have an initial state probability in a channel of
c=1, . . . , C₁,π_1,0^k,c;
a state transition probability from state sequence j to state:
_i_c_,a_1,i_C_|j;^k,c
and an observation probability:
b_t₀_,t₁^k,c(j_c).
In a continuous mixture with Gaussian components, the probability of the observation vector O is given by:

[0015]

\begin{matrix} b^{k, c} (j_{c}) = \sum_{m = 1}^{M_{j}^{k, c}} ω_{j, m}^{k, c} N (O, μ_{j, m}^{k, c}, U_{j, m}^{k, c}) \\ where μ_{j, m}^{k, c} and U_{j, m}^{k, c} \end{matrix}

are the mean and covariance matrix of the mth mixture of the Gaussian mixture corresponding to the jth state in the cth channel,
M_j^k,c
is the number of mixtures corresponding to the jth state of the cth channel, and
ω_j,m^k,c
is a weight associated with the corresponding mixture.

[0016] Observation sequences are used to form observation vectors later used in training and classifying. For example, the observation sequence for a two-dimensional image maybe formed from image blocks of size L_x×L_ythat are extracted by scanning the image from left-to-right and top-to-bottom. Adjacent image blocks may be designed to have an overlap by P_yrows in the vertical direction and P_xcolumns in the horizontal direction. In one possible embodiment, with blocks size of L_y=8 rows and L_x=8 columns, a six DCT coefficients (a 3×2 low-frequency array) may be employed to create the overlap.

[0017] The resulting array of observation vectors may correspond to size of T₀×T₁, where T₀and T₁are the number of observation vectors extracted along the height (H) and the width (W) of the image, respectively. T₀and T₁may be computed accordingly as:

[0018]

\begin{matrix} T_{0} = \frac{H - L_{y}}{L_{y} - P_{y}} + 1, \\ T_{1} = \frac{W - L_{x}}{L_{x} - P_{x}} + 1 \end{matrix}

[0019] Consecutive horizontal and vertical observation vectors may also be grouped together to form observation blocks. This may be used as a way to consolidate local observations and at the same time to reduce the total amount of observations. In practice, this data grouping serves application needs and improve recognition efficiency.

[0020] To compute the number of observation blocks, denote the number of observation blocks in the vertical and horizontal direction be T₀⁰and T₁⁰, respectively. Then,

[0021]

\begin{matrix} T_{0}^{0} = 1 \\ T_{1}^{0} = \frac{T_{1}}{C_{1}} \end{matrix}

In addition, denote the number of observation vectors in the horizontal and vertical direction within each observation block by T₀¹and T₁¹, respectively, where

[0022]

\begin{matrix} T_{0}^{1} = T_{1} \\ T_{1}^{1} = C_{1} \end{matrix}

Furthermore, denote O_t₀_,t₁_,cas the t₁th observation vector corresponding to the cth channel within the observation block t₀.

[0023] Although any suitable state sequence segmentation can be used, a modified Viterbi algorithm for the HMM-CHMM is preferred. Application of this modified Viterbi algorithm determines the optimal state and super state segmentation of the observation sequence. The best super state probability for the observation block t₀given super state i of super channel s, is denoted as P_t₀(i). Corresponding optimal state and optimal state sequence β_t₀_,t₁_,c(i) may then be computed for each super observation. The following states are first initialized:
δ(i)=π_0,0(i)P_t₀(i)
ψ₀(i)=0
The following states are then recursively determined:

[0024]

\begin{matrix} δ_{t_{0}} (i) = \max_{j} {δ_{t_{0} - 1} (j) a_{0, i | j} P_{t_{0}} (i)} \\ ψ_{t_{0}} (i) = \arg \max_{j} {δ_{t_{0} - 1} (j) a_{0, i | j} P_{t_{0}} (i)} \end{matrix}

The termination condition is then computed:
P=max_i{δ_T₀(i)}
α_T₀=arg max_i{δ_T₀(i)}
Based on the computed termination condition, a backtracking operation is performed:
α_T₀=ψ_0,t+1(α_T₀₊₁)
q_t₀_,t₁_,c⁰=α_t₀
q_t₀_,t₁_,c¹=β_t₀_,t₁_,c(α_t₀)

[0025] FIG. 3 is a flow diagram 40 illustrating training of an embedded HMM-CHMM based on the Viterbi algorithm, according to embodiments of the present invention. To train an HMM-CHMM based on given training data, observation vectors are first extracted from the training data set and organized in observation blocks (module 42). These observation blocks are uniformly segmented (module 44), replaced by an optimal state segmentation algorithm (module 46), have their model parameters estimated (module 48), and observation likelihood determined (module 50). As will be appreciated, the training may be iterative, with each training data set used individually and iteratively to update model parameters until the observation likelihood computed is smaller than a specified threshold.

[0026] More specifically, the training data set may be segmented along a first dimension according to the number of super states, into a plurality of uniform segments each of which corresponding to a super state. Based on the uniform segmentation at the super layer, the observation vectors within each uniform segment may then be uniformly segmented according to the number of channels and number of states of each child CHMM.

[0027] The density function of each state (including both super states as well as child states) may be initialized before the training takes place. For example, if Gaussian mixture model is adopted for each state, Gaussian parameters for each of the mixture component may need to be initialized. Different approaches may be employed to achieve the initialization of model parameters. For example, one embodiment may be implemented where the observation sequence assigned to each channel c and state j, and super state k and super channel s may be assigned to M_j^k,cclusters using, for example, the K-means algorithm.

[0028] During the process of training, the original uniform segmentation is updated based on the optimal state segmentation using the Viterbi algorithm or other suitable algorithms. To update the density function of a state, particular relevant parameters to be updated may be determined prior to the update operation.

[0029] The selection of a Gaussian mixture component for each state j channel c and super state k is also required. One exemplary criterion to make the selection may correspond to assigning the observation to O_t₀_,t₁_,c^(r)from the rth training sample in the training set to the Gaussian component for which the Gaussian density function

[0030]

N (O_{t_{0}, t_{1}, c}^{(r)}; μ_{j, m}^{k, c}, U_{j, m}^{k, c})

is the highest.

[0031] The parameters are then estimated using, for example, an extension of the segmental K-means algorithm. In particular, the estimated transition probability

[0032]

a_{0, i_{c} | j}^{'}

between super states i_sand j may be obtained as follows:

[0033]

a_{0, i | j}^{'} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \in_{t_{0}}^{(r)} (i, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} \in_{t_{0}}^{(r)} (i, l)}

where ∈_t₀^(r)(i,l) may equal to one if a transition from super state l to the super state i occurs for the observation block (t₀) and zero otherwise. The estimated transition probabilities

[0034]

a_{1, i_{c} | j}^{' k, c}

from embedded state sequence j to the embedded state i_cin channel c of super state k may then be obtained as follows,

[0035]

a_{1, i_{c} | j}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} θ_{t_{0}, t_{1}}^{(r)} (k, c, i_{c}, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} θ_{t_{0}, t_{1}}^{(r)} (k, c, i_{c}, l)}

where θ_t₀_,t₁^(r)(s, k, c, i_c, l) may be one if in the observation block (t₀) from the rth training sample a transition from state sequence j to state i_cin channel c occurs for the observation to O_t₀_,t₁_,c^(r)and zero otherwise.

[0036] The parameters of the selected Gaussian mixture component may also be accordingly updated. The involved Gaussian parameters may include a mean vector

[0037]

μ_{j, m, a}^{' k, c}

covariance matrix

[0038]

U_{j, m}^{' k, c}

of the Gaussian mixture, and the mixture coefficients

[0039]

ω_{j, m}^{' k, c}

for mixture m of state j channel c and super state k. The updated Gaussian parameters may be obtained according to the following formulations:

[0040]

\begin{matrix} μ_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m) O_{t_{0}, t_{1}, c}^{(r)}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \\ U_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m) (O_{t_{0}, s, t_{1}, c}^{r} - μ_{j, m}^{' k, c}) {(O_{t_{0}, t_{1}, c}^{(r)} - μ_{j, m}^{' k, c})}^{T}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \\ ω_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t1}^{} \sum_{m = 1}^{M} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \end{matrix}

where ψ_t₀_,t₁^(r)(k, c, j, m) may equal to one if the observation O_t₀_,t₁_,c^(r)is assigned to super state k, state j in channel c and mixture component m, and zero otherwise.

[0041] The update of parameters based on a training sample may be carried out iteratively This may be necessary because the Viterbi algorithm may yield different optimal segmentation during each iteration before convergence. Between two consecutive iterations, if the difference of observation likelihood computed with the Viterbi algorithm is smaller than a specified threshold, the iteration may be terminated. The HMM-CHMM corresponds to a complexity of quadratic with respect to the number of states in the model. In addition, HMM-CHMM may be efficiently implemented in a parallel fashion.

[0042] An alternative logical structure that includes an embedded CHMM-HMM (in contrast to an HMM-CHMM) is generically illustrated by FIG. 4. As seen in that Figure, a logical structure 60 of an embedded hidden Markov model-coupled hidden Markov model. As seen in FIG. 4, the CHMM-HMM is a hierarchical statistical model that includes a CHMM parent layer 62 (collectively formed from nodes 63) and a HMM child layer 64 (collectively formed from nodes 65). The child layer 64 associates one HMM node 65 to each node 63 in the parent layer 62, and the parameters of the individual HMMs remain independent from each other. Instead, the parameters of each child layer HMM depend upon the state of the connected parent node 63. Typically, for multidimensional data sets, the CHMM in the parent layer 62 is associated with at least one dimension, and the HMM child layers are associated with data in an orthogonal dimension with respect to the parent layer.

[0043] Formally defined, the elements of an embedded CHMM-HMM have: an initial super state probability π_0,0^sin super channel s and a super state transition probability from super state sequence j to super state i in super channel s, α_0,i|j^swhere super state refers to the state of the parent layer 32 CHMM node 33.

[0044] For each super state k the the super channel s the parameters of the corresponding HMM are defined so that the initial state probability is
π_1,0^s,k
the state transition probability from state j to state i is
a_1,i|j^s,k,c
and the observation probability is:
b_t₀_,t₁^s,k(j).
In a continuous mixture with Gaussian components, the probability of the observation vector O is given by:

[0045]

\begin{matrix} b^{s, k} (j) = \sum_{m = 1}^{M_{j}^{S, C}} ω_{j, m}^{s, k} N (O, μ_{j, m}^{s, k}, U_{j, m}^{s, k}) \\ where μ_{j, m}^{s, k} and U_{j, m}^{s, k} \end{matrix}

are the mean and covariance matrix of the mth mixture of the Gaussian mixture corresponding to the jth state and mth mixture,
M_j^s,k
is the number of mixtures corresponding to the jth state and,
ω_j,m^s,k
is a weight associated with the corresponding mixture.

[0046] Observation sequences are used to form observation vectors later used in training and classifying. For example, the observation sequence for a two-dimensional image may be formed from image blocks of size L_x×L_ythat are extracted by scanning the image from left-to-right and top-to-bottom. Adjacent image blocks may be designed to have an overlap by P_yrows in the vertical direction and P_xcolumns in the horizontal direction. In one possible embodiment, with blocks size of L_y=8 rows and L_x=8 columns, a six DCT coefficients (a 3×2 low-frequency array) may be employed to create the overlap.

[0047] The resulting array of observation vectors may correspond to size of T₀×T₁, where T₀and T₁are the number of observation vectors extracted along the height (H) and the width (W) of the image, respectively.

[0048] Consecutive horizontal and vertical observation vectors may also be grouped together to form observation blocks. This maybe used as a way to consolidate local observations and at the same time to reduce the total amount of observations. In practice, this data grouping serves application needs and improve recognition efficiency.

[0049] To compute the number of observation blocks, denote the number of observation blocks in the vertical and horizontal direction be T₀⁰and T₁⁰, respectively. Then,
T₀^0=C₀
T₁^0=T₁

[0050] In addition, denote the number of observation vectors in the horizontal and vertical direction within each observation block by T₀¹and T₁¹, respectively, where

[0051]

T_{0}^{1} = \frac{T_{1}}{C_{0}}

T_{1}^{1} = 1

Furthermore, denote O_t₀_,s,t₁as the t₁th observation vector corresponding to the observation block (t₀,s).

[0052] Although any suitable state sequence segmentation can be used, a modified Viterbi algorithm for the HMM-CHMM is preferred. Application of this modified Viterbi algorithm determines the optimal state and super state segmentation of the observation sequence. The best super state probability for the observation block (t₀,s) given super state i_sof super channel s, is denoted as P_t₀_s(i_s). Corresponding optimal state and optimal state sequence β_t₀_,t_1,s(i_s) may then be computed for each super observation. The following states are first initiazed:

[0053]

\begin{matrix} δ_{0, 0} (i) = \underset{s}{Π} π_{0, 0}^{s} (i_{s}) P_{t_{0}, s} (i_{s}) \\ ψ_{0, 0} (i) = 0 \end{matrix}

The following states are then recursively determined:

[0054]

\begin{matrix} δ_{0, t_{0}} (i) = \max_{j} {δ_{0, t_{0} - 1} (j) \prod_{s}^{} a_{0, i_{s}, | j_{s} - 1, j_{s}, j_{s} + 1}^{s} P_{t_{0}, s} (i_{s})} \\ ψ_{0, t_{0}} (i) = \arg \max_{j} {δ_{0, t_{0} - 1} (j) \prod_{s}^{} a_{0, i_{s} | j_{s} - 1, j_{s}, j_{s} + 1}^{s} P_{t_{0}, s} (i_{s})} \end{matrix}

The termination condition is then computed:
P=max_i{δ_T₀(i)}
{α_T₀_,1, . . . , α_T₀_,S}=arg max_i{δ_T₀(i)}
Based on the computed termination condition, a backtracking operation is performed:
{α_T₀_,1, . . . , α_T₀_,S}=ψ_0,t+1(α_T₀_+1,1, . . . , α_T₀_+1,S)
q_t₀_,s,t₁⁰=α_t₀_,s
q_t₀_,s,t₁¹=β_t₀_,s,t₁(α_t₀_,s)
Training of an embedded CHMM-HMM based on the Viterbi algorithm is substantially similar to that illustrated with respect to training of a HMM-CHMM as seen in FIG. 3. To train an CHMM-HMM based on given training data, observation vectors are first extracted from the training data set and organized in observation blocks. These observation blocks are segmented uniformly and at consecutive iterations through an optimal state segmentation algorithm, have their model parameters estimated and observation likelihood determined. As will be appreciated, the training may be iterative, with each training data set used individually and iteratively to update model parameters until the observation likelihood computed is smaller than a specified threshold.

[0055] More specifically, the training data set maybe segmented along a first dimension into S super channels. Then within each of such super channels, training data may further be uniformly segmented, according to the number of super states in each super channel, into a plurality of uniform segments each of which corresponding to a super state. Based on the uniform segmentation at the super layer, the observation vectors within each uniform segment may then be uniformly segmented according to the number of states of each child HMM. The density function of each state (including both super states as well as child states) may be initialized before the training takes place. For example, if Gaussian mixture model is adopted for each state, Gaussian parameters for each of the mixture component may need to be initialized. Different approaches may be employed to achieve the initialization of model parameters. For example, one embodiment may be implemented where the observation sequence assigned to state j, super state k and super channel s may be assigned to M_j^s,kclusters using, for example, the K-means algorithm.

[0056] During the process of training, the original uniform segmentation is updated based on the optimal state segmentation using the Viterbi algorithm or other suitable algorithms. To update the density function of a state, particular relevant parameters to be updated may be determined prior to the update operation. Depending on the density function used for each state, the selection may be carried out accordingly.

[0057] The selection of a Gaussian mixture components for each state j and super state k and super channel s is also required. One exemplary criterion to make the selection may correspond to assigning the observation O_t₀_,s,t₁^(r)from the rth training sample in the training set to the Gaussian component for which the Gaussian density function

[0058]

N (O_{t_{0}, s, t_{1}}^{(r)}; μ_{j, m}^{s, k}, U_{j, m}^{s, k})

is the highest.

[0059] The parameters are then estimated using, for example, an extension of the segmental K-means algorithm. In particular, the estimated transition probability a′_0,i_s_|j^sbetween super states i_sand the super state sequence j may be obtained as follows:

[0060]

a_{0, i_{s} | j}^{' s} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \in_{t_{0}}^{(r)} (s, i_{s}, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} \in_{t_{0}}^{(r)} (s, i_{s}, l)}

where ∈_t₀^(r)(s, i_s, l) may equal to one if a transition from state sequence l to the super state i_sin super channel s occurs for the observation block (t₀, s) and zero otherwise. The estimated transition probabilities between embedded states a′_l,i|j^s,kmay then be obtained as follows,

[0061]

a_{1, i | j}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} θ_{t_{0}, t_{1}}^{(r)} (s, k, i, l)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} θ_{t_{0}, t_{1}}^{(r)} (s, k, i, l)}

where θ_t₀_,t₁^(r)(s, k, i, l) may be one if in the observation block (t₀, s) a transition from state j to state is i_schannel c occurs for the observation O_t₀_,s,t₁^(r)and zero otherwise. The parameters of the selected Gaussian mixture component may also be accordingly updated. The involved Gaussian parameters may include a mean vector μ′_j,m^s,k, a covariance matrix U′_j,m^s,kof the Gaussian mixture, and the mixture coefficients ω′_j,m^s,kfor mixture m of state j in super state k and super channel s. The updated Gaussian parameters may be obtained according to the following formulations:

[0062]

\begin{matrix} μ_{j, m}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m) \times O_{t_{0}, s, t_{1}}^{(r)}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \\ U_{j, m}^{' s, k} = \frac{\begin{matrix} \sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m) \times \\ (O_{t_{0}, s, t_{1}}^{r} - μ_{j, m}^{' s, k}) \times {(O_{t_{0}, s, t_{1}}^{(r)} - μ_{j, m}^{' s, k})}^{T} \end{matrix}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \\ ω_{j, m}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t1}^{} \sum_{m = 1}^{M} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \end{matrix}

where ψ_t₀_,t₁^(r)(s, k, j, m) may equal to one if the observation O_t₀_,s,t₁^(r)is assigned to super state k in super channel s, state j and mixture component m, and zero otherwise.

[0063] The update of parameters based on a training sample maybe carried out iteratively. This may be necessary because the Viterbi algorithm may yield different optimal segmentation during each iteration before convergence. Between two consecutive iterations, if the difference of observation likelihood computed with the Viterbi algorithm is smaller than a specified threshold, the iteration may be terminated. The CHMM-HMM corresponds to a complexity of quadratic with respect to the number of states in the model. In addition, CHMM-HMM may be efficiently implemented in a parallel fashion.

[0064] As will be understood, reference in this specification to “an embodiment,” “one embodiment,” “some embodiments,” or “other embodiments” means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least some embodiments, but not necessarily all embodiments, of the invention. The various appearances “an embodiment,” “one embodiment,” or “some embodiments” are not necessarily all referring to the same embodiments.

[0065] If the specification states a component, feature, structure, or characteristic “may”, “might”, or “could” be included, that particular component, feature, structure, or characteristic is not required to be included. If the specification or claim refers to “a” or “an” element, that does not mean there is only one of the element. If the specification or claims refer to “an additional” element, that does not preclude there being more than one of the additional element.

[0066] Those skilled in the art having the benefit of this disclosure will appreciate that many other variations from the foregoing description and drawings may be made within the scope of the present invention. Accordingly, it is the following claims, including any amendments thereto, that define the scope of the invention.

Claims

The invention claimed is:

1. A pattern recognition method, comprising:

forming a hierarchical statistical model using a hidden Markov model (HMM) and a coupled hidden Markov model (CHMM), the hierarchical statistical model supporting a parent layer having multiple supernodes and a child layer having multiple nodes associated with each supernode of the parent layer; wherein either the parent layer is formed of an HMM and the child layer is formed of a CHMM, or the parent layer is formed of a CHMM and the child layer is formed of an HMM; the hierarchical statistical model applied to two dimensional data, with the parent layer describing data in a first direction and the child layer describing data in a second direction orthogonal to the first direction;

training the hierarchical statistical model using observation vectors extracted from a data set;

obtaining an observation vector sequence from a pattern to be recognized; and

identifying the pattern by finding a substantially optimal state sequence segmentation for the hierarchical statistical model.

2. The method according to claim 1, wherein the hierarchical statistical model defines an initial super state probability in a super channel.

3. The method according to claim 1, wherein the hierarchical statistical model defines super state transition probability from a sequence of states in the super channel.

4. The method according to claim 1, wherein the hierarchical statistical model defines observation likelihood given a state of a channel.

5. The method according to claim 1, wherein identifying the object comprises using a Viterbi algorithm to determine a substantially optimal state sequence for the hierarchical statistical model based on the observation vector sequence.

6. The method according to claim 1, wherein the data set includes two-dimensional data.

7. An article comprising a computer readable storage medium having stored thereon instructions that when executed by a machine result in:

training the hierarchical statistical model using observation vectors extracted from a data set;

obtaining an observation vector sequence from a pattern to be recognized; and

identifying the pattern by finding a substantially optimal state sequence segmentation for the hierarchical statistical model.

8. The article of claim 7, wherein the hierarchical statistical model defines an initial super state probability in a super channel.

9. The article of claim 7, wherein the hierarchical statistical model defines super state transition probability from a sequence of states in the super channel.

10. The article of claim 7, wherein the hierarchical statistical model defines observation likelihood given a state of a channel.

11. The article of claim 7, wherein identifying the object comprises using a Viterbi algorithm to determine a substantially optimal state sequence for the hierarchical statistical model based on the observation vector sequence.

12. The article of claim 7, wherein the data set includes two-dimensional data.

13. A system comprising:

a hierarchical statistical model to use both hidden Markov models and coupled hidden Markov models to model patterns, the hierarchical statistical model supporting a parent layer having multiple supernodes and a child layer having multiple nodes associated with each supernode of the parent layer; wherein either the parent layer is formed of a hidden Markov model (HMM) and the child layer is formed of a coupled HMM (CHMM), or the parent layer is formed of a CHMM and the child layer is formed of an HMM; the hierarchical statistical model applied to two dimensional data, with the parent layer describing data in a first direction and the child layer describing data in a second direction orthogonal to the first direction;

a training module to train for the hierarchical statistical model using observation vectors extracted from a data set; and

an identification module to obtain an observation vector sequence for a pattern to be recognized, and to identify the pattern by finding a substantially optimal state sequence segmentation for the hierarchical statistical model.

14. The system according to claim 13, wherein the hierarchical statistical model defines an initial super state probability in a super channel.

15. The system according to claim 13, wherein the hierarchical statistical model defines super state transition probability from a sequence of states in the super channel.

16. The system according to claim 13, wherein the hierarchical statistical model defines observation likelihood given a state of a channel.

17. The system according to claim 13, wherein the identification module is adapted to use a Viterbi algorithm to determine a substantially optimal state sequence for the hierarchical statistical model based on the observation vector sequence.

18. The system according to claim 13, wherein the data set includes two-dimensional data.

Patent Citations (44)

Patent	Date	Inventor	Cited By
US5454043(A)	1995-09-01	Freeman	Applicant
US5596362(A)	1997-01-01	Zhou	Applicant
US5710590(A)	1998-01-01	Ichige et al.	Applicant
US5754695(A)	1998-05-01	Kuo et al.	Applicant
US5850470(A)	1998-12-01	Kung et al.	Applicant
US5887069(A)	1999-03-01	Sakou et al.	Applicant
US6024852(A)	2000-02-01	Tamura et al.	Applicant
US6072494(A)	2000-06-01	Nguyen	Applicant
US6075895(A)	2000-06-01	Qiao et al.	Applicant
US6108005(A)	2000-08-01	Starks et al.	Applicant
US6128003(A)	2000-10-01	Smith et al.	Applicant
US6184926(B1)	2001-02-01	Khosravi et al.	Applicant
US6185529(B1)	2001-02-01	Chen et al.	Applicant
US6191773(B1)	2001-02-01	Maruno et al.	Applicant
US6212510(B1)	2001-04-01	Brand	Applicant
US6215890(B1)	2001-04-01	Matsuo et al.	Applicant
US6219639(B1)	2001-04-01	Bakis et al.	Applicant
US6222465(B1)	2001-04-01	Kumar et al.	Applicant
US6304674(B1)	2001-10-01	Cass et al.	Applicant
US6335977(B1)	2002-01-01	Kage	Applicant
US6385331(B2)	2002-05-01	Harakawa et al.	Applicant
US6594629(B1)	2003-07-01	Basu et al.	Applicant
US6609093(B1)	2003-08-01	Gopinath et al.	Applicant
US6624833(B1)	2003-09-01	Kumar et al.	Applicant
US6633844(B1)	2003-10-01	Verma et al.	Applicant
US6678415(B1)	2004-01-01	Popat et al.	Applicant
US6751354(B2)	2004-06-01	Foote et al.	Applicant
US6816836(B2)	2004-11-01	Basu et al.	Applicant
US6952687(B2)	2005-10-01	Andersen et al.	Examiner
US2002/0036617(A1)	2002-03-01	Pryor	Applicant
US2002/0093666(A1)	2002-07-01	Foote et al.	Applicant
US2002/0102010(A1)	2002-08-01	Liu et al.	Applicant
US2002/0135618(A1)	2002-09-01	Maes et al.	Applicant
US2002/0140718(A1)	2002-10-01	Yan et al.	Applicant
US2002/0161582(A1)	2002-10-01	Basson et al.	Applicant
US2003/0123754(A1)	2003-07-01	Toyama	Applicant
US2003/0144844(A1)	2003-07-01	Colmenarez et al.	Applicant
US2003/0154084(A1)	2003-08-01	Li et al.	Applicant
US2003/0171932(A1)	2003-09-01	Juang et al.	Applicant
US2003/0190076(A1)	2003-10-01	DeLean	Applicant
RU2112273(C1)	1995-08-01		Applicant
RU2093890(C1)	1997-10-01		Applicant
WO/36845	2000-06-01		Applicant
WO3/009218(A1)	2003-01-01		Applicant

Non-Patent Literature (34)

Matthew Brand, “Coupled Hidden Markov Models for Modeling Interacting Processes,” Learning and Common Sense Technical Report 405, Jun. 3, 1997, pp. 1-28, MIT Media Lab Perceptual Computing, USA.Applicant
A. Nefian and M. Hayes III, “An Embedded HMM-Based Approach for Face Detection and Recognition,” Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing: Mar. 15-19, 1999, IEEE, pp. 3553-3556, USA.Applicant
A. Nefian et al., “A Coupled HMM for Audio-Visual Speech Recognition,” 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing, IEEE, vol. 3 of 4, May 13, 2002, pp. 2013-2016, USA.Applicant
Ara V. Nefian, “Embedded Bayesian Networks for Face Recognition,” IEEE International Conference on Multimedia and Expo, IEEE, vol. 2, Aug. 26, 2002, pp. 133-136, USA.Applicant
Hennecke, et al: Automatic Speech Recognition System Using Acoustic and Visual Signals, IEEE, 1996.Applicant
Dupont et al: Audio-Visual Speech Modeling for Continuous Speech Recognition, Sep. 2000, IEEE Transactions on Multimedia, vol. 2, No. 3, pp. 141-151.Applicant
Potamianos et al: An Image Transform Approach for HMM Based Automatic Lipreading, Proc. Int. conf. Image Processing, 1998.Applicant
Potamianos et al: Linear Discriminant Analysis for Speechreading; IEEE Workshop on Multimedia Processing, Dec. 1998.Applicant
Chan: HHH-Based Audio-Visual Speech Recognition Integrating Geometric and Appearance-Based Visual Features, IEEE 2001.Applicant
Pavlovic: Dynamic Bayesian Networks for Information Fusion with Applications to Human-Computer Interfaces; Thesis, University of Urbana-Champaign, 1999, pp. iii-ix and 63-81.Applicant
Rezek, et al: Coupled Hidden Markov Models for Biosignal Interaction; Advances in Medical Signal and Information Processing, Sep. 4-6, 2000; pp. 54-59.Applicant
Fu, et al: Audio-Visual Speaker Identification Using Coupled Hidden Markov Models; 2003 Int'l Conference on Image Processing (ICIP), Sep. 14-17, 2003; vol. 2, pp. 29-32.Applicant
Kristjansson, et al: Event-Coupled Hidden Markov Models; 2000 IEEE Int'l Conference on Multimedia and Expo, Jul. 30-Aug. 2, 2000; vol. 1; pp. 385-388.Applicant
Pavlovic: Multimodal Tracking and Classification of Audio-Visual Features; 1998 Int'l Conference on Image Processing, ICIP Proceedings; Oct. 4-7, 1998, vol. 1; pp. 343-347.Applicant
Wikipedia, definition of Hidden Markov Model, 3 pages.Applicant
Wikipedia, definition of Viterbi Algorithm, 5 pages.Applicant
Rezek, et al: Learning Interaction Dynamics with Coupled Hidden Markov Models; IEEE Proceedings—Science, Measurement and Technology, Nov. 2000; vol. 147, Issue 6; pp. 345-350.Applicant
Logan et al: Factorial Hidden Markov Models for Speech Recognition: Preliminary Experiments; Cambridge Research Laboratory; Technical report Series; CRL 97/7; Sep. 1997.Applicant
Dugad: Tutorial on Hidden Markov Models; Technical Report No.: SPANN-96, May 1996, pp. 1-16.Applicant
Kennedy, et al: Identification of Coupled Markov Chain Model with Application; Proceedings of the 31st IEEE Conference on Decision and Control, Dec. 16-18, 1992; vol. 4, pp. 3529-3534.Applicant
Ramesh, et al: Automatic Selection of Tuning Parameters for Feature Extraction Sequences; Proceedings IEEE Computer Society Conference on Computer vision and Pattern Recognition; Jun. 21-23, 1994, pp. 672-677.Applicant
Liang, et al: Speaker Independent Audio-Visual Continuous Speech Recognition; Aug. 2002; Multimedia and Expo, vol. 2, pp. 25-28; IEEE.Applicant
Luettin et al.: Asynchronous Stream Modelling for Large Vocabulary Audio-Visual Speech Recognition, Proceedings of the 2001 IEEE Int'l Conference of Acoustics, Speech and Signal Processing (ICASSP'01), May 7-11, 2001, pp. 169-172.Applicant
Gordan: A Temporal Network for Support Vector Machine Classifiers for the Recognition of Visual Speech, Methods and Applications of Artificial Intelligence: Proceedings of the 2nd hellenic Conference on AI (SETN 2002), Thessaloniki, Greece, Apr. 11-12, 2002, pp. 355-365.Applicant
Ming-Husan Yang et al.: Detecting Faces in Images: A Survey; IEEE trans Pattern Analysis and Machine Intelligence, vol. 24, No. 1, Jan. 2002, pp. 34-58.Applicant
Yongmin Li et al.: Multi-view Face Detection Using Support Vector Machines and Eigenspace Modelling, Proceedings on the Int'l Conference on Knowledge-based Intelligent Engineering Systems and.Applicant
Batra: Modeling and Efficient Optimization for Object-Based Scalability and Some Related Problems, IEEE Transactions onImage processing, vol. 9, No. 10, Oct. 10, 2000, pp. 1677-1692.Applicant
Pending U.S. Appl. No. 10/326,368; Office Action dated Jul. 25, 2006.Applicant
Int'l Application No.: PCT/US03/31454 Written Opinion dated Oct. 12, 2006.Applicant
Brand: Coupled Hidden Markov Models for Modeling Interacting Processes; MIT Media Lab Perceptual Computing/Learning and Common Sense Technical Report 405; Nov. 3, 1996, revised Jun. 3, 1997; XP-002253550; 28 pages.Applicant
NETI et al.: Large-Vocabulary Audio-Visual Speech Recognition: A Summary of the Johns Hopkins Summer 2000 Workshop.Applicant
Pending U.S. Appl. No. 10/143,459, filed May 9, 2002, inventor: Lu Hong Liang; Final Office Action dated Oct. 31, 2006.Applicant
Wikipedia: Definition of Linear Discriminant Analysis.Applicant
Pending U.S. Appl. No.: 10/143,459 filed May 9, 2002, inventor: Lu Hong Liang; Final Office Action dated Oct. 31, 2006.Applicant

Abstract

Description

Field of the Invention

[0001] The present invention relates to computer mediated pattern detection. More particularly, the present invention relates to improved Bayesian networks for classifying data.

Background

Brief Description of the Drawings

[0005] FIG. 1 schematically illustrates a data classification system;

[0006] FIG. 2 generically illustrates an embedded hidden Markov model-coupled hidden Markov model (HMM-CHMM) structure;

[0007] FIG. 3 is a flow diagram illustrating training of an embedded HMM-CHMM; and

[0008] FIG. 4 generically illustrates an embedded coupled hidden Markov model-hidden Markov model (CHMM-HMM) structure.

Detailed Description

[0015]

\begin{matrix} b^{k, c} (j_{c}) = \sum_{m = 1}^{M_{j}^{k, c}} ω_{j, m}^{k, c} N (O, μ_{j, m}^{k, c}, U_{j, m}^{k, c}) \\ where μ_{j, m}^{k, c} and U_{j, m}^{k, c} \end{matrix}

[0018]

\begin{matrix} T_{0} = \frac{H - L_{y}}{L_{y} - P_{y}} + 1, \\ T_{1} = \frac{W - L_{x}}{L_{x} - P_{x}} + 1 \end{matrix}

[0020] To compute the number of observation blocks, denote the number of observation blocks in the vertical and horizontal direction be T₀⁰and T₁⁰, respectively. Then,

[0021]

\begin{matrix} T_{0}^{0} = 1 \\ T_{1}^{0} = \frac{T_{1}}{C_{1}} \end{matrix}

In addition, denote the number of observation vectors in the horizontal and vertical direction within each observation block by T₀¹and T₁¹, respectively, where

[0022]

\begin{matrix} T_{0}^{1} = T_{1} \\ T_{1}^{1} = C_{1} \end{matrix}

Furthermore, denote O_t₀_,t₁_,cas the t₁th observation vector corresponding to the cth channel within the observation block t₀.

[0024]

\begin{matrix} δ_{t_{0}} (i) = \max_{j} {δ_{t_{0} - 1} (j) a_{0, i | j} P_{t_{0}} (i)} \\ ψ_{t_{0}} (i) = \arg \max_{j} {δ_{t_{0} - 1} (j) a_{0, i | j} P_{t_{0}} (i)} \end{matrix}

[0030]

N (O_{t_{0}, t_{1}, c}^{(r)}; μ_{j, m}^{k, c}, U_{j, m}^{k, c})

is the highest.

[0031] The parameters are then estimated using, for example, an extension of the segmental K-means algorithm. In particular, the estimated transition probability

[0032]

a_{0, i_{c} | j}^{'}

between super states i_sand j may be obtained as follows:

[0033]

a_{0, i | j}^{'} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \in_{t_{0}}^{(r)} (i, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} \in_{t_{0}}^{(r)} (i, l)}

where ∈_t₀^(r)(i,l) may equal to one if a transition from super state l to the super state i occurs for the observation block (t₀) and zero otherwise. The estimated transition probabilities

[0034]

a_{1, i_{c} | j}^{' k, c}

from embedded state sequence j to the embedded state i_cin channel c of super state k may then be obtained as follows,

[0035]

a_{1, i_{c} | j}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} θ_{t_{0}, t_{1}}^{(r)} (k, c, i_{c}, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} θ_{t_{0}, t_{1}}^{(r)} (k, c, i_{c}, l)}

[0036] The parameters of the selected Gaussian mixture component may also be accordingly updated. The involved Gaussian parameters may include a mean vector

[0037]

μ_{j, m, a}^{' k, c}

covariance matrix

[0038]

U_{j, m}^{' k, c}

of the Gaussian mixture, and the mixture coefficients

[0039]

ω_{j, m}^{' k, c}

for mixture m of state j channel c and super state k. The updated Gaussian parameters may be obtained according to the following formulations:

[0040]

\begin{matrix} μ_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m) O_{t_{0}, t_{1}, c}^{(r)}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \\ U_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m) (O_{t_{0}, s, t_{1}, c}^{r} - μ_{j, m}^{' k, c}) {(O_{t_{0}, t_{1}, c}^{(r)} - μ_{j, m}^{' k, c})}^{T}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \\ ω_{j, m}^{' k, c} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t1}^{} \sum_{m = 1}^{M} ψ_{t_{0}, t_{1}}^{(r)} (k, c, j, m)} \end{matrix}

where ψ_t₀_,t₁^(r)(k, c, j, m) may equal to one if the observation O_t₀_,t₁_,c^(r)is assigned to super state k, state j in channel c and mixture component m, and zero otherwise.

[0045]

\begin{matrix} b^{s, k} (j) = \sum_{m = 1}^{M_{j}^{S, C}} ω_{j, m}^{s, k} N (O, μ_{j, m}^{s, k}, U_{j, m}^{s, k}) \\ where μ_{j, m}^{s, k} and U_{j, m}^{s, k} \end{matrix}

[0049] To compute the number of observation blocks, denote the number of observation blocks in the vertical and horizontal direction be T₀⁰and T₁⁰, respectively. Then,
T₀^0=C₀
T₁^0=T₁

[0050] In addition, denote the number of observation vectors in the horizontal and vertical direction within each observation block by T₀¹and T₁¹, respectively, where

[0051]

T_{0}^{1} = \frac{T_{1}}{C_{0}}

T_{1}^{1} = 1

Furthermore, denote O_t₀_,s,t₁as the t₁th observation vector corresponding to the observation block (t₀,s).

[0053]

\begin{matrix} δ_{0, 0} (i) = \underset{s}{Π} π_{0, 0}^{s} (i_{s}) P_{t_{0}, s} (i_{s}) \\ ψ_{0, 0} (i) = 0 \end{matrix}

The following states are then recursively determined:

[0054]

\begin{matrix} δ_{0, t_{0}} (i) = \max_{j} {δ_{0, t_{0} - 1} (j) \prod_{s}^{} a_{0, i_{s}, | j_{s} - 1, j_{s}, j_{s} + 1}^{s} P_{t_{0}, s} (i_{s})} \\ ψ_{0, t_{0}} (i) = \arg \max_{j} {δ_{0, t_{0} - 1} (j) \prod_{s}^{} a_{0, i_{s} | j_{s} - 1, j_{s}, j_{s} + 1}^{s} P_{t_{0}, s} (i_{s})} \end{matrix}

[0058]

N (O_{t_{0}, s, t_{1}}^{(r)}; μ_{j, m}^{s, k}, U_{j, m}^{s, k})

is the highest.

[0060]

a_{0, i_{s} | j}^{' s} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \in_{t_{0}}^{(r)} (s, i_{s}, j)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} \in_{t_{0}}^{(r)} (s, i_{s}, l)}

[0061]

a_{1, i | j}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} θ_{t_{0}, t_{1}}^{(r)} (s, k, i, l)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} \sum_{l}^{} θ_{t_{0}, t_{1}}^{(r)} (s, k, i, l)}

[0062]

\begin{matrix} μ_{j, m}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m) \times O_{t_{0}, s, t_{1}}^{(r)}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \\ U_{j, m}^{' s, k} = \frac{\begin{matrix} \sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m) \times \\ (O_{t_{0}, s, t_{1}}^{r} - μ_{j, m}^{' s, k}) \times {(O_{t_{0}, s, t_{1}}^{(r)} - μ_{j, m}^{' s, k})}^{T} \end{matrix}}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \\ ω_{j, m}^{' s, k} = \frac{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t_{1}}^{} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)}{\sum_{r}^{} \sum_{t_{0}}^{} \sum_{t1}^{} \sum_{m = 1}^{M} ψ_{t_{0}, t_{1}}^{(r)} (s, k, j, m)} \end{matrix}