Cognitive Computational Model Using Machine Learning Algorithm in Artificial Intelligence Environment

First, we consider the algorithm function. The DBN algorithm belongs to a kind of neural network in machine learning and can have both unsupervised learning and supervised learning functions [16]. DBN is a probability generation model, which establishes a combination distribution between observation data and tags. By training the weights between units, the neural network can train the data information with the maximum probability as the goal. DBN is used for feature recognition, data classification and probability generation. DBN algorithm has strong practicability, wide application fields and high expansion performance. It can be used in handwriting recognition, speech recognition and image processing in machine learning [17].

Second, we look at the structure and principle of the algorithm. The components of DBN are multilayer units (visible units and hidden units). Visible units can be used to receive input, and hidden units can be used for feature extraction [18]. The connection between the top two layers of units has no direction, forming the associative memory. The other neural layers belong to directed connection, and the bottom layer belongs to data vectors. A unit represents a dimension of the data vector. Further, the restricted Boltzmann machine (RBM) is a component of DBN, which can train the DBN layer by layer. The data vector can be used to judge the hidden layer in each layer. Then, this layer is used as the data vector of the unit in the next higher layer. The RBM can function independently as a clusterer. There are only two layers of units in RBM, one of which is the visible layer for training input data. The other layer is the hidden layer, which plays the function of feature detector. Figure 1 shows its structure.

Fig. 1

In Figure 1, the units (grey boxes) on the left form the hidden layer, and the units on the right form the visible layer. Each layer can be represented by a vector; each unit represents a dimension, and bidirectional connection is carried out between two layers.

In order to make the research content more general, each layer of the DBN algorithm training is a Markov random field model, and each model contains a visible layer and a hidden layer. U = (U₁,...,U_n) represents the variable matrix in the visible layer, P = (P₁,P_n) represents the variable matrix of the hidden layer. Based on (U, P) Gibbs distribution, the variable matrix P of the hidden layer is used to describe the marginal distribution of the variable matrix U of the visible layer. (1) S(U)=∑mS(U,P)=1K∑mexp⁡(−∈(U,P)) S(U)=\sum_{m} S(U, P)=\frac{1}{K} \sum_{\mathrm{m}} \exp (-\in(U, P))

In Eq. (1), K=∑U,Pexp⁡(−∈(U,P))K=\sum_{U, P} \exp (-\in(U, P)). Therefore, the likelihood function of parameter λ can be defined as follows: (2) ln⁡L(λ∣U)=ln⁡S(λ∣U)=ln⁡∑Pexp⁡(−∈(U,P))−ln⁡∑U,Pexp⁡(−∈(U,P)) \ln L(\lambda \mid U)=\ln S(\lambda \mid U)=\ln \sum_{P} \exp (-\in(U, P))-\ln \sum_{U, P} \exp (-\in(U, P))

According to Eq. (2), the gradient of parameter attribute adjustment is calculated. (3) αln⁡L(λ∣U)αλ=−ln⁡∑PP(P∣U)∂∈(U,P)αλ+ln⁡∑PP(P,U)∂∈(U,P)αλ \frac{\alpha \ln L(\lambda \mid U)}{\alpha \lambda}=-\ln \sum_{P} P(P \mid U) \frac{\partial \in(U, P)}{\alpha \lambda}+\ln \sum_{P} P(P, U) \frac{\partial \in(U, P)}{\alpha \lambda}

Here, α represents the parameter adjustment value.

Each RBM contains a hidden layer and a visible layer, which represent different variables. The visible and hidden layers are not connected with itself. If the values are 0 and 1, the energy equation of the RBM can be obtained. (4) E(U,P)=−∑x=1a∑y=1bUxOxyPy−∑x=1bIxUx−∑y=1aRyPy. <math display="block"><mrow><mi>E</mi><mo stretchy="false">(</mo><mi>U</mi><mo>,</mo><mi>P</mi><mo stretchy="false">)</mo><mo>=</mo><mo>&#x2212;</mo><munderover><mrow><mo>&#x2211;</mo></mrow><mrow><mi>x</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>a</mi></mrow></munderover><munderover><mrow><mo>&#x2211;</mo></mrow><mrow><mi>y</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>b</mi></mrow></munderover><msub><mrow><mi>U</mi></mrow><mrow><mi>x</mi></mrow></msub><msub><mrow><mi>O</mi></mrow><mrow><mi>x</mi><mi>y</mi></mrow></msub><msub><mrow><mi>P</mi></mrow><mrow><mi>y</mi></mrow></msub><mo>&#x2212;</mo><munderover><mrow><mo>&#x2211;</mo></mrow><mrow><mi>x</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>b</mi></mrow></munderover><msub><mrow><mi>I</mi></mrow><mrow><mi>x</mi></mrow></msub><msub><mrow><mi>U</mi></mrow><mrow><mi>x</mi></mrow></msub><mo>&#x2212;</mo><munderover><mrow><mo>&#x2211;</mo></mrow><mrow><mi>y</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>a</mi></mrow></munderover><msub><mrow><mi>R</mi></mrow><mrow><mi>y</mi></mrow></msub><msub><mrow><mi>P</mi></mrow><mrow><mi>y</mi></mrow></msub><mo>.</mo></mrow></math>

Here, O, I and R represent the paranoid coefficients, and a and b represent the maximum ranges of the paranoid coefficient.

According to Eq. (4), it can be identified that in the energy equation, the weight should be adjusted globally and the bias coefficient should be relatively independent. If RBM is regarded as an image, the image only contains the connection between the hidden layer and the visible layer, and there is no connection inside the layer. According to Gibbs distribution and Eq. (4), the conditional probability equation can be obtained. (5) p(Uy=1∣P)=σ(∑y=1aOxyPy+Iy)p(Py=1∣U)=σ(∑x=1aOxyUy+Rx) \begin{aligned} &p\left(U_{y}=1 \mid P\right)=\sigma\left(\sum_{y=1}^{a} O_{x y} P_{y}+I_{y}\right) \\ &p\left(P_{y}=1 \mid U\right)=\sigma\left(\sum_{x=1}^{a} O_{x y} U_{y}+R_{x}\right) \end{aligned}

where p is the probability and σ is the coefficient.

Then, the RBM network is trained using the contrastive divergence algorithm. After the parameter of paranoid ideation is adjusted, the learning function of parameter λ can be obtained. (6) λT+1=ϕ⋅λT+μ⋅{ααλ[1V∑Z=1VlnL⁡L(λ∣Uy)]−β⋅λT} \lambda^{T+1}=\varphi \cdot \lambda^{T}+\mu \cdot\left\{\frac{\alpha}{\alpha \lambda}\left[\frac{1}{V} \sum_{Z=1}^{V} \ln L\left(\lambda \mid U_{y}\right)\right]-\beta \cdot \lambda^{T}\right\}

where φ represents inertia factor, µ represents the learning rate, and β represents the penalty factor.

The RBM training process aims to calculate the probability set that can generate many training samples. In this probability set, the training samples have the largest probability. The decisive factor of the probability combination is the weight size, so the purpose of RBM training is to find the optimal weight.

Third, we move on DBN training. DBN is a neural network based on RBM, which can be either a generation model or a discriminant model. The process of training the DBN involves the use of unsupervised layer-by-layer analysis to get the weights. Thus, the training process of DBN should be carried out layer by layer. In each layer, the hidden layer prediction should be carried out according to the data vector. Then, the hidden layer is used as the information vector for adjacent higher layers. In other words, many RBMs can be combined to form a DBN. The visible layer (input value) of the next RBM is the hidden layer (output value) of the previous RBM. During the training process, the RBM of the previous layer must be fully trained before the RBM of the next layer can be trained, which goes all the way to the end layer. Finally, the units are the essential components of the neural network, and the DBN is composed of many units and RBMs. The restricted layer of the DBN network structure comprises the visible layer and the hidden layer, and there are corresponding connections between layers. However, there is no connection between the neurons in the interior of the layer (grey boxes). By training the units of the hidden layer, the association between the higher-order data displayed in the visible layer can be mined. Figure 2 shows the composition of a DBN network.

Fig. 2

Fourth, we come to the essence of the DBN algorithm. From the perspective of unsupervised learning, the goal is to preserve the nature of the original features to the greatest extent and to reduce the dimensions of features. From the perspective of supervised learning, the goal is to keep the error rate of classification as low as possible. Irrespective of whether supervised or unsupervised learning is adopted, the DBN algorithm is essentially a process of feature learning, in other words, a process to get the best feature expression.

Fifth, we detail the DBN training process. The first RBM needs to be fully trained; its weight and offset are fixed, and its hidden units are used as the import vector of the next RBM; then, after the second RBM is fully trained, the second RBM is overlapped with the first RBM. It is necessary to repeat the above steps several times. If the data in the training set contain labels, when the top RBM is trained, the visible layers in the RBM must contain both visible units and the units representing classification labels. Both of them are trained at the same time, as shown in Figure 3 below.

Fig. 3

For each training data, the corresponding labelled units can be set to 1 after being opened, and the remaining ones will be closed and set to 0. The parts labelled M₁ and M₂ in Figure 3 are the top labels for training in RBM.

The neural unit at the top of the network evaluates and classifies tasks based on the data generated by the lower network. When the DBN is trained at the end of the network layer, it can use the feedforward neural network to slightly adjust the previous evaluation and classification according to the data information with labels. However, the processing method of this study is better than the direct use of a feedforward neural network, because the DBN has higher efficiency, and it only changes the weight parameters and then trains in a small area. Moreover, the training speed is very fast and the convergence time is relatively short.