5.1.1.1 Supervised Learning

In supervised learning, the training set is provided with a set of labels to predict the target value. Its main application includes regression (target values are continuous values) such as predicting weather and population growth. Additionally, it can also be applied to classification problems like speech recognition, fraud detection, digit recognition, diagnostics, etc., by employing algorithms like support vector machine (SVM), random forest, and others. Learning here is divided into two parts. (i) Training phase: Here the data is provided with a label, and the algorithm learns the relationship between input data and labels and determines the output for the test data. (ii) Testing phase: Here the data without a label is provided as input, and the algorithm finds the relationship between input data and previous labels and determines the output for the test data [7].

5.1.1.2 Unsupervised Learning

In unsupervised learning training, a set is not provided with a set of labels. This is meant for finding useful structures or knowledge in data when no labels are available. It is also used in detecting abnormal server access patterns, detecting irregular access to websites, outlier detection, social network analysis, organizing computing clusters, etc. It solves clustering problems like target marketing, customer segmentation, and recommendation system. It also tries to find patterns to predict future values and also applied for dimensionality reduction problems.

5.1.1.3 Semi‐Supervised Learning

It is a combination of supervised and unsupervised learning, that is, both labeled and unlabeled data are used. This is an improvement over unsupervised learning [8].

5.1.1.4 Reinforcement Learning

In reinforcement learning, appropriate steps or a sequence of steps are to be taken for a given situation to maximize the payoff. Its applications include AI gaming, real‐time decisions, skill acquisition, and robot navigation. Figure 5.2 explains broad categories of learning.

5.1.2.1 Basic Terminology

• Representation: understand the problem
• Evaluation: find effectiveness
• Optimization: solve the problem and feature representation.

5.1.2.2 Training Process

The training set consists of N training samples labeled as {( x_i, y_i), …( x_m, y_m)}. ( x_i, y_i) is a training instance. x_i is the ith m‐dimensional input vector, y_i is the ith n‐dimensional output vector. x = {x₁, x₂, . . , x_m} is the m‐dimensional input matrix and y = {y₁, y₂, . . , y_n} is the n‐dimensional output matrix. Artificial Intelligence techniques build a model from the input and map its corresponding output until the system stabilizes.

5.1.3.1 Generative

In the case of Generative models, individual class distribution is modeled. The Generative models are based on joint probability distribution p(x, y) = p(y) * p(y ∣ x) and conclude based on Bayes rule and select the most likely label. When conditional independence is not convinced, they prove less accurate. They need fewer data to train and test well when working with missing data. They are applied to find hidden parameters and are used for solving classification problems. Generative models have less chance to overfit but are affected by outliers and need more computation. Examples of the Generative model include naïve Bayes and linear discriminative analysis.

5.1.3.2 Discriminative

Discriminative models are based on conditional probability p(Y ∣ X = x). These models learn the boundary between classes, so they are useful for classification. It does not work well when we have missing data, and needs sufficient data to train these models. These models are not affected by outliers, but they tend to over fit and require less computation than generative models. Examples of discriminative models include logistic regression, SVMs, etc.

5.2.1.1 Training Boltzmann Machine

Training of Boltzmann machine is done through contrastive divergence (CD) [11] using Gibbs Sampling:

1. The visible state is fed with the training dataset.
2. Positive phase:

   Hidden state nodes are calculated from visible nodes known as positive statistics of edge E_ij which is P(H_j = 1|V). The individual activation probabilities for the hidden layer can be calculated as

   

   
   

3. Negative phase:

   In the negative phase, the input data is reconstructed using the hidden state values known as negative statistics of edge E_ij, which is P(V_i = 1|H).

   The individual activation probabilities for visible layer nodes can be calculated as:

   
   
4. Update the weight of the edge:

   The updated weight is calculated as:

   

   and α is the learning rate.

5. Repeat till the threshold is achieved.

5.2.1.2 Applications of RBM

RBM has various applications in the field of classification of data in the form of text, images, etc. It can also be used for reducing a number of variables for obtaining principle variables known as dimensionality reduction and can be used for feature selection and feature extraction. Other applications include processing of unstructured data, images, videos, collaborative filtering, topic modeling, missing value issue, etc. RBM has been applied for supervised and unsupervised learning, and an energy‐efficient RBM processor has been proposed for on‐chip learning for IoT [12]. Sourav Bhattacharya and Nicholas D. Lane [13] proposed RBM as a DL architecture for robust activity recognition using smart watches, like sleeping patterns and emotional states. They are used for predicting energy consumption in IoT.

The variants of RBMs include (i) discriminative restricted Boltzmann machine (DRBM): they are meant for the selection of parameters, which are analytical for performance. (ii) Hybrid restricted Boltzmann machine (HRBM): it includes features of both generative and discriminative models. (iii) Conditional restricted Boltzmann machine (CRBM). (iv) Robust Boltzmann machine (RoBM): it deals with eliminating noise and influence of corrupted pixels in an image.

5.2.2.1 Training DBN

1. Train the first layer of RBM using CD algorithm and learn the weights W_1.
2. Freeze the weight vector W₁ of the first layer. Use the H₁ (output of the first layer) from P (H₁|V) as input for the next layer.
3. Freeze the weights of second layer W₂ that defines the second layer features. Use the H₂ (output of the second layer) as input to the next layer.
4. Proceed recursively for the next layers.

5.2.2.2 Applications of DBN

It has been used to avoid overfitting and underfitting problems. It has been used for recognizing and generating images, capturing motion in a video, dimensionality reduction, audio classification, and natural language classification. They can act as feature detectors and has been applied for real‐life applications like electroencephalography [17] and drug discovery [18]. Hinton et al. [16] have deployed for the recognition and classification of handwritten characters. Ni Gao et al. [19] have deployed DBN in intrusion detection. It can provide new ideas for further research in Intrusion Detection Systems. They are suitable for hierarchical feature extraction as they use a greedy layer‐by‐layer training algorithm. It has been used for threat identification in the security and classification of faults in IoT.

5.2.3.1 Training of AE

AE consist of three main parts: (i) encoder, (ii) code, and (iii) decoder. Figure 5.5 explains the architecture of AE. AE are self‐supervised learning algorithms because they generate their labels. The loss function measures how close the reconstructed output is to the input and is given by mean squared error:

AE add a bottleneck to the network. The bottleneck forces the network to create the compress of the input. It works well if the data are correlated and does not work well if the data are not correlated.

Encoder: converts input x into some hidden representation h(x):

Decoder: reconstructs the input :

where b, c are biases and w is the weight vector.

w ^* = w^T and are known as tied weights.

Variants of AE include: (i) denoising autoencoders [20]: these are the type of AE where we try to reconstruct the input from corrupted input data. They try to undo the effects of corruption in data. They improve the accuracy of OCR. (ii) Sparse autoencoder and (iii) variational autoencoder.

Variational autoencoder is a standard type of AE which shares the basic architecture of AE. Here, training is done to avoid overfitting. Encoding is done over a distribution of latent space (z). Low‐dimensional view of high‐dimensional input has been created.

5.2.3.2 Applications of AE

The applications of AE include anomaly detection, information retrieval, data denoising, etc. They are more powerful than principal component analysis (PCA). Mahmood Yousefi‐Azar et al. [21] proposed AE‐based feature learning for internet security and for various classifiers for anomaly detection in publically available datasets. Changjie Hu et al. [22] proposed improved version of AE for dimensionality reduction. It has the same number of input and output and is well suited for dimensionality reduction and feature extraction. In IoT, it can be used for emotion recognition and fault diagnosis in machinery.

5.2.4.1 Layers of CNN

• Convolution layer: This is the main layer in a CNN. It acts as a filter to an input; when this filter is applied continuously to the input, we get feature map. Feature map gives us the locations of the important information in an image. CNN can learn filters specific to a particular training dataset. Here, multiplication is performed between input and a set of weights (called filter or kernel). The multiplication between the kernel and input is performed elementwise called dot product. The multiplication is followed by addition. Kernel size is smaller than input size; the same kernel has been applied multiple times to the different portions of an input by systematically overlapping different portions of input from left to right and top to bottom.

  Whenever we are applying convolution, our output shrinks. So to avoid shrinking, we can use padding to get output dimensions the same as an input. Convolutions are of two types: (i) valid and (ii) same. In case of valid, the output we get is smaller than input and no padding is applied. In the other type, the output dimensions are the same as input, and padding is applied here. The three main parameters that control the size of output are: (i) stride, (ii) kernel size, and (iii) padding.

  The step size of the filter when processing the convolution is known as a stride. The same stride size is applied both horizontally and vertically. More the stride values are smaller, the output we get. Kernel size is the filter that we apply to the input. If we want to output the same as input, then we need to include padding. CNN has the property that a part of output depends on only a smaller part of an input that is known as the sparsity of connections. A part of the image is possibly important for other parts of an image known as parameter sharing.

• Pooling layer: The function of a pooling layer is to reduce the representation size of an image. It works on feature maps extracted from the previous layer. It gives us the representation of a summary of the input, so that processing and computation are reduced. The common pooling functions include
  1. Max pooling: It calculates the maximum value of a portion of a feature map.
  2. Average pooling: It calculates the average of a portion of a feature map.
  3. Global pooling: This is also known as a global pooling layer. It downsamples whole feature map to a single unit.
• Fully connected layer: It is the last layer of CNN. Here, we get the class label for our input. It has a lot of connections forming a complete graph with the output from the previous layer. It gives us the decision of whole processing. Input to this layer is fed in the form of a flattened vector. At this layer, results are available.

  Variants of CNN: CNN has become popular in the last few years. It has been applied to various fields and has achieved the best performance. Popular variants of CNN include (i) LeNet, (ii) AlexNet, (iii) ResNet, (iv) VGG, (v) GoogLeNet, (vi) Inception Networks, (vii) Xception, etc.

5.2.4.2 Activation Functions Used in CNN

Activation functions are nonlinear functions that decide output for the next layer. The activation functions used in CNN include:

1. Sigmoid Function: It is also known as logistic activation function. It translates input (−∞, +∞) to (0, 1) used for binary class problem.
   

   Figure 5.7 shows the graph of sigmoid function. Soft‐max logistic equation is used for multi‐class problems. This is used in the last layer of CNN.

2. Tanh(): This is same as sigmoid activation function but here it translates to (−1, 1). Figure 5.8 shows a graph of Tanh() function.
   
3. ReLU: ReLU stands for a rectified linear unit. It gives linear output for positive values and 0 for negative values. It converges faster and is used in the convolution layer of CNN in the middle layers.
   

Figure 5.9 shows a graph of ReLU function. It has the problem of dying ReLU, which means if it gets 0 value, it is difficult for it to get back. Its variants include Leaky ReLU, PReLU, and ELU.

5.2.4.3 Applications of CNN

CNN has proven satisfying in the field of image recognition, image transformations, face recognition, speech recognition, image labeling, recommendation system, handwriting recognition, natural language processing (NLP), behavior recognition, object recognition, etc. Iman Azimi et al. [23] have deployed CNN in the IoT healthcare system. Alberto Pacheco [24] et al. deployed it for the development of Smart Classroom and IoT Computing. A wavelet‐based CNN approach has been proposed for automated machinery fault diagnosis [25]. Parkinson's disease patients' prediction can be done using CNNs [26]. CNNs are also used for mapping sequence of ECG samples to a sequence of rhythm classes to develop high diagnostic performance similar to professionals [27]. CNNs have been used for automatic segmentation of MRI scans to predict the risk of osteoarthritis [28]. V. Gulshan et al. [27] have used CNNs to identify retinopathy. A CNN‐based approach is introduced for surface integration inspection in [29] to extract patch features and find defected areas via thresholding and segmenting. A wavelet‐based CNN approach has been proposed for automated machinery fault diagnosis [25].

5.2.5.1 Training of GANs

It mainly consists of two neural network architectures that contest with one another for learning. The two networks compete with each other to generate some new probability distribution that is indistinguishable to original. GANs generate the imitate data from some original data. It consists of two main parts:

• Discriminator: The role of the discriminator module is to discriminate between original and fraud data.
• Generator: The role of the generator module is to generate the fraud data in such a way that it can deceive the discriminator. That is, data produced here must look like the original.

Figure 5.10 shows the basic framework of GANs for identifying fake or real things. Learning in GANs means deceiving the discriminator, i.e. the discriminator treats fake instances as real. The discriminator weights are kept fixed, the error is backpropagated, and weight adjustment is done at the generator. Figure 5.11 explains the basic architecture of GANs.

5.2.5.2 Variants of GANs

The variants of GANs include progressive GANs, conditional GANs, Image to Image GANs, and Cycle GANs.

5.2.5.3 Applications of GANs

GANs have been applied mainly in the field of data generation [30]. They have been successfully applied in the field of computer vision and NLP. The applications in the field of computer vision include image translation, image synthesis, and video generation. Ledig et al. [31] have applied GANs for improvement in the resolution of an image. Li and Wand [32] applied it in the field of texture synthesis. Tulyakov et al. [33] proposed GANs for video generation. They have a great opportunity to be applied in new fields and have been applied for the labeling of unlabeled images [34].

5.2.6.1 Training of RNN

Training RNN is comparable to training a Neural Network. Here also backpropagation algorithm is used, but with modification. The parameters are shared by all time‐steps in the network; the gradient at each output depends on the estimates of the current time step, and previous time steps. To calculate the gradient at each point, backpropagation is needed to be called backpropagation through time (BPTT). Figure 5.12 explains the basic architecture of RNN.

The hidden state at a particular time is a function (F_w) of previous state (S_t−1) and current input (X_t).

where W_s, W_x are weights.

5.2.6.2 Applications of RNN

The RNN is trained using BPTT algorithm [35]. It is difficult to learn long‐term dependencies in the input sequence due to the decrease in error gradient as it propagates backward known as vanishing gradient problem. This leads to larger weights which in turn lead to unstable training. B. Chandra, Rajesh Kumar Sharma [36] has proposed the improved version of RNN which reduced vanishing gradient problem and has promising results for image classification. It learns complex patterns in the IoT.

5.2.7.1 Training of LSTM

LSTM consists of a chain of repeating modules called memory cells which consists of four gates. These four gates decide what information to keep or discard, handle dependencies, and handle the storage of information [38]. The four gates are:

• Input gate (i)
• Output gate (o)
• Forget gate (f)
• Memory gate (c).

The information is passed based on the weights and these weights are adjusted during the learning process. Each memory cell correlates with a time step [39]. It removes information through the forget gate, it specifically controls information transmission using the sigmoid function. Thus, outputs values between 0 and 1. One determines to keep the information and zero to discard the information. They maintain a constant error throughout; therefore, they are able to learn over multiple layers and time steps. They sustain information for a more prolonged time. Each cell has its individual decision power related to information. Forget gate removes irrelevant information, thus helps in intensification in the process. Input gate adds information, output gate selects necessary information for the output. Figure 5.13 explains the architecture of LSTM.

5.2.7.2 Applications of LSTM

LSTM has mainly been developed for handling sequential data, NLP [40]. They have been developed for intrusion detection for learning patterns [41]. It has also been proposed for the classification of voltage dips and automatic extraction of features [42]. LSTM has become one of the most popular techniques in DL; research is going on in this field. It has been also proposed to control dynamic systems [43]. It has proven better for the classification and identification of fault with 99.80% accuracy [44]. The other applications include text generation, handwriting recognition [45], handwriting generation, music generation, language translation, and image captioning. In industries, it has been used for the analysis of equipments of IoT [46] and also for time series data generation in IoT, anomaly detection, etc. Time series data are highly complex, dynamic data, and has higher dimensionality, so it requires IoT of resources. LSTM allows forgetting unnecessary data and thus reduces the computation power needed. It can keep track of dependencies between data.