Fractional-Order Deep Backpropagation Neural Network

In recent years, the research of artificial neural networks based on fractional calculus has attracted much attention. In this paper, we proposed a fractional-order deep backpropagation (BP) neural network model with L2 regularization. The proposed network was optimized by the fractional gradient descent method with Caputo derivative. We also illustrated the necessary conditions for the convergence of the proposed network. The influence of L2 regularization on the convergence was analyzed with the fractional-order variational method. The experiments have been performed on the MNIST dataset to demonstrate that the proposed network was deterministically convergent and can effectively avoid overfitting.


Introduction
It is well known that artificial neural networks (ANNs) are the abstraction, simplification, and simulation of the human brains and reflect the basic characteristics of the human brains [1]. In recent years, great progress has been made in the research of deep neural networks. Due to the powerful ability of complex nonlinear mapping, many practical problems have been successfully solved by ANNs in the fields of pattern recognition, intelligent robot, automatic control, prediction, biology, medicine, economics, and other fields [2,3]. BP neural network is one of the most basic and typical multilayer forward neural networks, which are trained by backpropagation (BP) algorithm. BP, which is an efficient way for optimization of ANNs, was firstly introduced by Werbos in 1974. Then, Rumelhart and McCelland et al. implemented the BP algorithm in detail in 1987 and applied it to the multilayer network version of Minsky [4][5][6].
The fractional calculus has a history as long as the integral order calculus. In the past three hundred years, the theory of fractional calculus has made great progresses [7][8][9][10][11]. Its basics are differentiation and integration of arbitrary fractional order. Nowadays, fractional calculus is widely used in diffusion processes [12][13][14], viscoelasticity theory [15], automation control [16][17][18], signal processing [19][20][21], image processing [22][23][24][25], medical imaging [26][27][28], neural networks [29][30][31][32][33][34][35][36][37], and many other fields. Due to the long-term memory, nonlocality, and weak singularity characteristics [29][30][31][32][33][34][35][36][37], fractional calculus has been successfully applied to ANNs. For instance, Boroomand constructed the Hopfield neural networks based on fractional calculus [37]. Kaslik analyzed the stability of Hopfield neural networks [30]. Pu proposed a fractional steepest descent approach and offered a detailed analysis of its learning conditions, stability, and convergence [38]. Wang applied the fractional steepest descent algorithm to train BP neural networks and proved the monotonicity and convergence of a three-layer example [33]. However, there are three limitations in the proposed fractional-order BP neural network models in [33]. First, the neural network in [33] just had 3 layers, which was actually a shadow network and was not proper to demonstrate its potential for deep learning. Second, the fractional order V of this model was restricted to (0, 1] without reasonable analysis. Third, the loss function did not contain the regularization term, which is an efficient way to avoid overfitting, especially when the training set has small scalar. Overfitting means that the model has high prediction accuracy on training set but has the low prediction accuracy on testing set. This makes the generalization ability of the model poor, and the application value is greatly reduced.
In this paper, we proposed a deep fractional-order BP neural network with 2 regularization term, and the fractional-order V could be any positive real number. With the fractional-order variational method, the influence of 2 regularization on the convergence of the proposed model was exploited. The performance of the proposed model was evaluated on the MINST dataset.

Computational Intelligence and Neuroscience
The structure of the paper is as follows: in Section 2, the definitions and simple properties of fractional calculus are introduced. In Section 3, the proposed fractional-order multilayer BP neural networks are given in detail. In Section 4, the necessary conditions and the influence of 2 regularization for the convergence of the proposed BP algorithm are stated. In Section 5, experimental results are presented to illustrate the effectiveness of our model. Finally, the paper is concluded in Section 6.

Background Theory for Fractional Calculus
In this section, the basic knowledge of fractional calculus is introduced, including the definitions and several simple properties used in this paper.
The following is the G-L definition of fractional derivative: where − V denotes the fractional differential operator based on G-L definition, ( ) denotes a differintegrable function, V is the fractional order, [ , ] is the domain of ( ), Γ is the Gamma function, and [⋅] is the rounding function.
The R-L definition of fractional derivative is as follows: where − V denotes the fractional differential operator based on G-L definition; = [V + 1]. Moreover, the G-L fractional derivative can be deduced from the definition of the R-L fractional derivative.
The Caputo definition of fractional derivative is as follows: where V is the fractional differential operator based on Caputo definition, = [V + 1].
Fractional calculus is more difficult to compute than integer calculus. Several mathematical properties used in this paper are given here. The fractional differential of a linear combination of differintegral functions is as follows: where ( ) and ( ) are differintegral functions and and are constants. The fractional differential of constant function ( ) = C, (C is a constant) is different under different definitions: For the G-L definition, For the R-L definition, And for the Caputo definition According to (6), (7) and (8), we can know that for the G-L and R-L definition, the fractional differential of constant function is not equal to 0. Only with the Caputo definition, the fractional differential of constant function equals to 0, which is consistent to the integer-order calculus. Therefore, the Caputo definition is widely used in solving engineering problems and it was employed to calculate the fractionalorder derivative in this paper. The fractional differential of function ( ) = ( − ) , ( > −1) is as follows:

Fractional-Order Deep BP Neural Networks.
In this section, we introduce the fractional-order deep BP neural network with L layers. , = 1, 2, . . . , L, is the number of neurons for the -th layer. W = ( ) +1 × denotes the weight matrix connecting the -th layer and the ( + 1)-th layer. denotes the corresponding activation function for the -th layer. and are the input and the corresponding ideal output of the -th sample and the training sample set is { , } =1 .
= ( 1 , 2 . . . , +1 ) denotes the total inputs of -th layer. If neurons in the -th layer are not connected to any neurons in previous layer, these neurons are called external outputs of the -th layer, denoted as 1 . On the contrary, if neurons in the -th layer are connected to every neuron in previous layer, these neurons are called internal outputs ofth layer, denoted as 2 . = ( 1 , 2 . . . , ) denotes the total Computational Intelligence and Neuroscience 3 outputs of -th layer. The forward computing of the fractionalorder deep BP neural networks is as follows: Particularly, external outputs can exist in any layer except the last one. With the square error function, the error corresponding to -th sample can be denoted as: where denotes the -th element of , denotes the -th element of .
The total error of the neural networks is defined as In order to minimize the total error of the fractionalorder deep BP neural network, the weights are updated by the fractional gradient descent method with Caputo derivative. Let = 1, 2, ..., . The backpropagation of fractional-order deep BP neural networks can be derived with the following steps.
Firstly, we define that According to (13), we can know that Then the relationship between and +1 can be given by Then, according to the chain rule and (17), we have The updating formula is where ∈ N denotes the -th iteration and > 0 is the learning rate.

Fractional Deep BP Neural Networks with 2 Regularization.
Fractional-order BP neural network can be overfitted easily when the training set has small scalar. 2 regularization is a useful way to avoid models to be overfitted without modifying the architecture of network. Therefore, by introducing the 2 regularization term into the total error, the modified error function can be presented as where ‖ ‖ 2 denotes the sum of squares of all weights and ≥ 0 denotes the regularization parameter. By introducing (18), we have The updating formula is where ∈ N denotes the -th iteration and > 0 is the learning rate.

Convergence Analysis
In this section, the convergence of the proposed fractionalorder BP neural network is analyzed. According to previous studies [39][40][41][42], there are four necessary conditions for the convergence of BP neural networks: (1) The activation functions , ( = 1, 2, . . . , ) are bounded and infinitely differentiable on R and all of their corresponding derivatives are also continuous and bounded on . This condition can be easily satisfied because the most common sigmoid activation functions are uniformly bounded on and infinitely differentiable.
(2) The boundedness of the weight sequence {( ) } is valid during training procedure and is the domain of all weights with certain boundary.
(3) The learning rate > 0 has an upper bound.
(4) Let denote the weights matrix that consists of all weights and = { | V 2 = 0} be the V-order stationary point set of the error function. One necessary condition is that is a finite set.
Then, the influence of 2 regularization on the convergence is derived by using the fractional-order variational method.
According to (20), 2 is defined as a fractional-order multivariable function. The proposed fractional-order BP algorithm is to minimize 2 . Let denote the fractionalorder extreme point of 2 and denotes an admissible point. In addition, is composed of 1 , 2 , . . . , −1 where ( = 1, 2, . . . , − 1) denotes the weights matrix between the -th and ( + 1)-th layer when 2 reaches the extreme value. is composed of 1 , 2 , . . . , −1 where corresponds to . The initial weights are random values, so the initial points of weights can be represented as + ( − 1) , where is a vector that consists of small parameters 1 , 2 , . . . , −1 , and corresponds to and . If = 1, it means = 1( = 1, 2, . . . , − 1), then + ( − 1) = , and 2 reaches the extreme value. Thus, the process of training the BP neural networks from a random initial weight to U can be treated as the process of training with a random initial value to = 1.
The fractional-order derivative of 2 on + ( − 1) is given as where V is the fractional order, which is a positive real number.
From (23), we can see that when = 1, if the V-order differential of ( + ( − 1) ) with respect to is existent, 1 ( ) has a V-order extreme point and we have In this case, the output of each layer in the neural networks is still given by (10) and (11) and the input of each layer is turned into the following: When = 1, we have Without loss of generality, according to (18), for the -th layer of the networks, the V-order differential of with respect to can be calculated as where denotes the column vector 1 .
Since the value of is stochastic, according to variation principle [43], to allow (24) to be set up, a necessary condition is that for every layer of the networks Secondly, without loss of generality, for 2 ( ) we have To allow (29) to be set up, a necessary condition is With (28) and (30), the Euler-Lagrange equation of V 2 | =1 can be written as Equation (31) is the necessary condition for the convergence of the proposed fractional-order BP neural networks with 2 regularization. From (31), we can see that if > 0, then ( +1 ⋅( ) ) ̸ = 0. ( +1 ⋅( ) ) is the first-order derivative of in terms of and can be calculated by and input sample . It means that the extreme point U of the proposed algorithm is not equal to the extreme point of integer-order BP algorithm or fractional-order BP algorithm. changes with the different value of and V. In addition, it is also clear that the regularization parameter is bounded since the values of input samples and weights are bounded and V is a constant during the training process.

Experiments
In this section, the following simulations were carried out to evaluate the performance of the presented algorithm. The simulations have been performed on the MNIST handwritten digital dataset. Each digit in the dataset is a 28 × 28 image. Each image is associated with a label from 0 to 9. We divided each image into four parts, which were top-left, bottom-left, bottom-right, and top-right, and each part was a 14 × 14 matrix. We vectorized each part of the image as a 196 × 1 vector and each label as a 10 × 1 vector.
In order to identify the handwritten digits in MNIST dataset, a neural network with 8 layers was proposed. Figure 1 shows the topological structure of the neural networks.   the first layer are sigmoid functions, which can be given as follows: The MNIST dataset has a total number of 60000 training samples and 10000 testing samples. The simulations demonstrate the performance of the proposed fractionalorder BP neural network with 2 regularization, fractionalorder BP neural network, traditional BP neural network, and traditional BP neural network with 2 regularization. To evaluate the robustness of our proposed network for a small set of training samples, we set the number of training samples to be (10000, 20000, 30000, 40000, 50000, and 60000). Different fractional V-order derivatives were employed to compute the gradient of error function, where V = 1/9, 2/9, 3/9, 4/9, 5/9, 6/9, 7/9, 8/9, 9/9, 10/9, 11/9, 12/9, 13/9, 14/9, 15/9, 16/9, 17/9, 19/9, and 20/9 separately (V = 9/9 = 1 corresponds to standard integer-order derivative for the common BP; V ̸ = 2 because if V = 2 the change of weights after each iteration is 0, and the weights of the neural networks cannot be updated). The learning rate was set to be 3 and the batch size was set to be 100. The number of epochs was 300. Two main metrics-training accuracy and testing accuracy-were used to measure the performance of the results from different networks. Each network was trained 5 times and the average values were calculated.
In order to explore the relationship between the fractional orders and the neural network performance, the fractionalorder neural networks with different orders were trained. Figure 2 shows the results of different networks with different sizes of training set. We can find that when the fractional order V exceeds 1.6, both the training and testing accuracies declined rapidly, and when the fractional order V > 2, the performances of the fractional BP neural networks were much poorer than that with 0 < V < 2. The results of V = 19/9 and 20/9 were shown in Table 1 as examples. This result is consistent with that for describing physical problems, and usually the limitation 0 < V < 2 is adopted in the fractionalorder models.
From Figure 2, it can be observed that, with the increase of the size of training set, the performances of the networks were improved visibly. Furthermore, it is also obvious that the training and testing accuracies raised gradually with increasing fractional orders and then reached the peak while V equaled 10/9 or 11/9 order. After that, the training and testing accuracies began to decline rapidly. Table 2 shows the optimal orders under training set and testing set separately with different size of training set and it can be noticed that the optimal orders almost concentrated in 10/9 and 11/9. The only exception is that when the number of training samples was 50000, the training accuracy of order 1 was slightly higher than that in 10/9 or 11/9 order case. Generally, for the MNIST dataset the performances of fractional-order BP neural networks are better than integer order.
The performance of the proposed fractional-order BP neural networks with 2 regularization and the performance comparison with integer-order BP neural networks (IOBP), integer-order BP neural networks with 2 regularization, and fractional-order BP neural networks (FOBP) in terms of training and testing accuracy are shown in Table 3 and the change of the testing accuracy with the iterations was given in Figure 3 In Table 3 and Figure 3, it can be seen that, after the addition of 2 regularization to BP neural networks, the training accuracy is slightly decreased but the testing accuracy significantly increased, which indicated that adding 2 regularization can effectively suppress overfitting and improve the generalization of BP neural networks. Furthermore, it can be noticed that after adding 2 regularization the performance of fractional-order BP neural network is better than integer order. One important merit of the 2 regularization is that it gained more benefit while the training set is small. The most possible reason is that the network trained with the smallest number of training samples was affected most by the overfitting. With the increase of the training samples, the model gradually changed from overfitting to underfitting, so the improvement of the regularization method became faint.
Then, the stability and convergence of the proposed fractional-order BP neural networks with 2 regularization  are demonstrated in Figures 4 and 5. We used the network with optimal order, which means that the size of training set was 60000, fractional-order V was 11/9, and the regularization parameter was 3 × 10 −6 . Figure 4 shows the change of the total error 2 during the training process. Without loss of generality, the change of V 5 20,20 2 was randomly selected and Figure 5 shows the change of it during the training process. It is clear to see that 2 and V w 2 converged fast and stably and were finally close to zero. These observations effectively verify the proposed algorithm is deterministically convergent.

Conclusion
In this paper, we applied fractional calculus and regularization method to deep BP neural networks. Different from previous studies, the proposed model had no limitations on the number of layers and the fractional-order was extended to arbitrary real number bigger than 0. 2 regularization was also imposed into the errorless function. Meanwhile, we analyzed the profits introduced by the 2 regularization on the convergence of this proposed fractional-order BP network. The numerical results support that the fractional-order BP neural networks with 2 regularization are deterministically convergent and can effectively avoid the overfitting phenomenon. Then, how to apply fractional calculus to other more complex artificial neural networks is an attracted topic in our future work.

Data Availability
The code of this work can be downloaded at https://github .com/BaoChunhui/Deep-fractional-BP-neural-networks.

Conflicts of Interest
The authors declare that they have no conflicts of interest.