Applied Learning Algorithms for Intelligent IoT

Pethuru Raj Chelliah, Usha Sakthivel, Susila Nagarajan, Pethuru Raj Chelliah, Usha Sakthivel, Susila Nagarajan

• Publisher: Auerbach Publications
• Year: 2021
• ISBN:
  9781000461367
• Edition: 1
• Topic: Informatica
• Subtopic: Intelligenza artificiale (IA) e semantica

This book vividly illustrates all the promising and potential machine learning (ML) and deep learning (DL) algorithms through a host of real-world and real-time business use cases. Machines and devices can be empowered to self-learn and exhibit intelligent behavior. Also, Big Data combined with real-time and runtime data can lead to personalized, prognostic, predictive, and prescriptive insights. This book examines the following topics:

• Cognitive machines and devices

• 
  
  

• Cyber physical systems (CPS)

• 
  
  

• The Internet of Things (IoT) and industrial use cases

• 
  
  

• Industry4.0 for smarter manufacturing

• 
  
  

• Predictive and prescriptive insights for smarter systems

• 
  
  

• Machine vision and intelligence

• 
  
  

• Natural interfaces

• 
  
  

• K-means clustering algorithm

• 
  
  

• Support vector machine (SVM) algorithm

• 
  
  

• A priori algorithms

• 
  
  

• Linear and logistic regression

• 
  
  

Applied Learning Algorithms for Intelligent IoT clearly articulates ML and DL algorithms that can be used to unearth predictive and prescriptive insights out of Big Data. Transforming raw data into information and relevant knowledge is gaining prominence with the availability of data processing and mining, analytics algorithms, platforms, frameworks, and other accelerators discussed in the book. Now, with the emergence of machine learning algorithms, the field of data analytics is bound to reach new heights.

This book will serve as a comprehensive guide for AI researchers, faculty members, and IT professionals. Every chapter will discuss one ML algorithm, its origin, challenges, and benefits, as well as a sample industry use case for explaining the algorithm in detail. The book's detailed and deeper dive into ML and DL algorithms using a practical use case can foster innovative research.

1

Convolutional Neural Network in Computer Vision

D. Aishwarya¹ and R.I. Minu²

¹Research Scholar, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India

²SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India

DOI: 10.1201/9781003119838-1

Contents

Introduction

Convolutional Neural Network (CNN)

Distinctive Properties of CNN

Activation Functions for CNN

Loss Function

Datasets and Errors

Bias and Variance

Overfitting and Underfitting

Understanding Padding and Stride

Padding

Stride

Parameters and Hyper Parameters

CONV Layer

Filter

Feature Map

Convolution Operation

Key Points about Convolution Layers and Filters

Pooling Layer

Key Points about Pooling Layer

Types of Pooling

Forward Propagation

Calculating the Parameters

Activation Shape and Size

Backward Propagation

Optimizers

Other Ways to Improve the Performance of CNN

Image Data Augmentation

Deeper Hidden Layers

Early Stopping

Application of CNN

Image Classification

RCNN and Object Detection

Region-Based Convolutional Neural Network (R-CNN)

Fast R-CNN

Faster R-CNN

You Only Look Once (YOLO)

Transfer Learning

Transfer Learning for CNN

Fine-Tuning or Freezing?

Neural Style Transfer (NST)

Introduction

Emergence of artificial intelligence (AI) has made machines “Smart.” By “Smart,” we mean the ability of the machine (possibly a computer) to process the given data with rationality, like an intelligent human being, and take a suitable decision. AI achieves this intelligence with the help of its subsets of algorithms (see Figure 1.1).

Figure 1.1 Subsets of Artificial Intelligence

AI is a broader field that includes machine learning (ML), artificial neural network (ANN), and deep learning (DL) as its subsets. AI is achieved by ANN, which is a network of artificial neurons resembling a biological neuron and the nervous system. ML enables the machines to process the data, make decision, and optimize the results based on experience. ML primarily requires a feature engineer to do the feature engineering process. Also, ML works its best for structured dataset (mostly comprises numerical data) but is proven inefficient for unstructured types.

To process the unstructured and continuous data like video, speech, etc., we need more complex neural network. Here comes in the DL with better results on processing the unstructured data. One of the biggest advantages of DL is that it features engineering automatically, that is, no manual feature selection is required. Also, ML requires manual intervention when the algorithm is struck with optimization problem, which is also automated in DL. DL plays the key role in implementing and visualizing the actual AI. Deep learning, however, requires a large amount of dataset for the purpose of training, which in turn requires a deep network of layers of neurons to process the features and hence the name “deep learning.”

DL comprises many algorithms for different types of data. Of those, convolutional neural network (CNN) is mainly used to process images and solve computer vision problems. The idea of CNNs can be traced back to 1980, when K. Fakushima proposed “neocognition” – a hierarchical multi-layered neural networks, which is self-organized and unaffected by the translation operation in the dataset (shift in positions of patterns). It was primarily created for pattern recognition.

The first standard CONVNet was introduced by Yann LeCunn in 1998, in his research article titled “Object Recognition with Gradient Based Learning.” It is important to note that Yann LeCunn has cited Fakushima’s article in his research work, and the CONVNet of LeCunn has a structure similar to neocognition with back propagation added to the network. Thenceforth, multiple improvements and the extension were made to it.

Convolutional Neural Network (CNN)

Convolutional neural network is primarily used for operations like image classification, segmentation, and other computer vision operations. CNN is also called shape/space invariant artificial neural network (SIAAN) based on its shared weights and translation invariance properties. The key idea behind CNN is to capture the local patterns through convolution operation.

In general, each CNN architecture has a common working flow. A CNN consists of multiple layers of convolution and pooling and a fully connected ANN. A layer in CNN may comprise one convolution layer and one pooling layer together as a single layer if the count of convolution layer and pooling layer is same. Most commonly used type is two convolution layers followed by a single pooling layer.

The number of layers in a CNN architecture depends on the complexity of the application. The output of the last layer is passed to an ANN. The convolution layers and the pooling layers always perform the same operation of learning the input image. The fully connected networks perform the required operations, such as regression, classification, etc.

At each convolution layer, a local pattern is obtained by applying filters throughout the image. Filters can also be mentioned as kernels. The Initial/lower-level layers of the CNN obtain the elementary pattern of the image ...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Contents
• Contributor List
• Chapter 1: Convolutional Neural Network in Computer Vision
• Chapter 2: Trends and Transition in the Machine Learning (ML) Space
• Chapter 3: Next-Generation IoT Use Cases across Industry Verticals Using Machine Learning Algorithms
• Chapter 4: A Panoramic View of Cyber Attack Detection and Prevention Using Machine Learning and Deep Learning Approaches
• Chapter 5: Regression Algorithms in Machine Learning
• Chapter 6: Machine Learning-Based Industrial Internet of Things (IIoT) and Its Applications
• Chapter 7: Employee Turnover Prediction Using Single Voting Model
• Chapter 8: A Novel Implementation of Sentiment Analysis Toward Data Science
• Chapter 9: Conspectus of k-Means Clustering Algorithm
• Chapter 10: Systematic Approach to Deal with Internal Fragmentation and Enhancing Memory Space during COVID-19
• Chapter 11: IoT Automated Spy Drone to Detect and Alert Illegal Drug Plants for Law Enforcement
• Chapter 12: Expounding k-Means-Inspired Network Partitioning Algorithm for SDN Controller Placement
• Chapter 13: An Intelligent Deep Learning-Based Wireless Underground Sensor System for IoT-Based Agricultural Application
• Chapter 14: Predicting Effectiveness of Solar Pond Heat Exchanger with LTES Containing CuO Nanoparticle Using Machine Learning
• Index