Up and Running Google AutoML and AI Platform

Building Machine Learning and NLP Models Using AutoML and AI Platform for Production Environment (English Edition)

Amit Agrawal

• Publisher: BPB Publications
• Year: 2020
• ISBN:
  9789388511926
• Topic: Ciencia de la computación
• Subtopic: Procesamiento del lenguaje natural

Learn how to work towards making the most out of a career in emerging tech Key Features

• Understand the core concepts related to careers in emerging tech.
• Learn innovative, exclusive, and exciting ways to design a successful career in ET.
• Reduce your learning curve by examining the career trajectories of eminent ET professionals.
• Ways to evolve and adapt to changing ET paradigms.
• Practical perspective from the field.
• 
  Description
  Cracking the emerging tech code will help you attain your Emerging Technology (ET) career goals faster without spending years in committing avoidable mistakes, recovering from them, and learning things the hard way. You can apply practical tips in areas such as improving your ability to craft market-friendly use cases and evolving a solution approach in new and diverse tech or business environments, to propel forward your career in strategic and proactive ways.It outlines ways in which you can explore and capitalize on hidden opportunities while working on important career aspects. The anecdotes and solutions provided will aid you in getting an inside out view to reduce your learning curve. This book will help you in gaining both magnitude and direction in your ET career journey and prevent you from getting overwhelmed or pinned down by the forces of ET.Authored by an ET professional, this book will take you through a series of steps to deepen your understanding of the forces that shape one's ET career and successfully dealing with them.It also helps bust myths, addresses fallacies, and common misconceptions that could harm one's career prospects. There are also practical and easy-to-adopt tips, methods, tracking mechanisms, and information that will improve career standing and professional growth.This book makes it easy for you to enhance your employability and job market relevance so that you can sprint towards a rewarding career. What will you learn
  Through this book, you will connect with ways and means to build a strong and rewarding emerging tech career. You will be able to work on identifying the right technology and employer, enhancing employability and differentiation in the job market, addressing challenges and connecting with enablers, accurate growth strategies and execution principles. Who this book is for
  This book is for current and aspiring emerging tech professionals, students, and anyone who wishes to understand ways to have a fulfilling career in emerging technologies such as AI, blockchain, cybersecurity, IoT, space tech, and more. Table of Contents
  1. Introduction2. The best ET for me and some myth bursting3. Getting prepared and charting a roadmap4. Identifying the requirements and getting help5. Dealing with headwinds and drawing a career change action plan6. Building an ET friendly résumé and finding the right employer7. Getting hired through social media8. Job search9. Impressing the emerging tech jury10. The secret sauce11. Becoming a thought leader12. Measuring success and making course corrections13. Drawing the two-year plan14. Building your leadership capabilities15. To start-up or not?16. Communications skills: getting it right17. Building a personal brand18. Post-script About the Authors
  Prayukth has been actively involved in productizing and promoting cross eco-system collaboration in the IoT space for over half-a-decade. In recent years, he has focused on exposing APT groups, global footprint, and in evaluating the evolving threat landscape surrounding IoT and OT environments.In his current role, he has taken Subex's IoT business to new geographies. Your Linkedin profile: https://www.linkedin.com/in/prayukthkv/
  

CHAPTER 1

Introduction to Artificial Intelligence

The digital revolution in the 1980s brought with it a wide spectrum of options, where machines or computers could ease out the life of humans in ways that had not been imagined earlier. Using machines for different tasks by programming them, made them problem-solvers in all sorts of domains that were earlier reserved for specific tasks. This ease of programmability also opened the doors for self-programming or rather Intelligent Systems and, hence, the idea of Artificial Intelligence (AI) got started.

With Artificial Intelligence, several human characteristics of “Thinking, Decision-Making, Trends’ Analysis, and Problem-Solving ” are being done by computer systems or integrated machines. Artificial Intelligence has now created a mark in almost every arena where smart or intelligent solutions can be integrated. Machine Learning and Artificial Intelligence algorithms are helping computing machines to adopt to human-like intelligence and decision-making skills in tasks like medical diagnosis, proving mathematical theorems, search engines, and voice recognition.

Structure

In this chapter, we will cover the following topics:

• Machine Learning
  • Supervised Learning
  • Unsupervised Learning
  • Reinforcement Learning
  • Applications of Machine Learning
• Natural Language Processing
  • Pre-processing
  • Parts of Speech (POS) Tagging
  • Entity Extraction
  • Sentiment Analysis
  • Document Summarization
  • Topic Modeling

Objectives

After studying this chapter, you should be able to:

• Understand the basics of Machine Learning and Natural Language Processing
• Understand the applications of Machine Learning and Natural Language Processing

Machine Learning

Machine Learning can be easily defined as a branch of Artificial Intelligence that deals with self-training or programing and adapting according to the given data that speaks of the problem statement. The problem statements can range from mere identification of entities to a tag or class to decision-making if the given input favors an expected output. Some algorithms can help one identify if a fruit is Tomato or Persimmon, which are pretty similar looking; such algorithms classify an entity in different classes. Another example can be predicting the stock market trends based on historical behaviors and real-time variables.

The major difference from traditional computer software is that the developer does not write the code that instructs the system the way to tell the difference between the two objects. Instead, an algorithm has been taught the way to generate these decisions by being trained on an oversized amount of data. One thing that has to be understood is that if any Machine Learning Algorithm is a solution for a problem statement, a relevant and considerably sized dataset is always a prerequisite to start with.

Machine learning is categorized into three main types, namely, Supervised Learning, Unsupervised Learning, and Semi-Supervised Learning. We are going to look at an overview of these categories in this chapter.

Supervised Learning

Supervised Learning is a term used for algorithms where we have training data with accurate labels against them. The significance of such algorithms is the accuracy of labeled data or prediction as input helps in having crisp decision boundaries with the training of algorithm. The labeled data provides the guidance or knowledge to the system to understand and map the data. The labeling of the data is done by human subject matter experts.

Another advantage with having a labeled dataset is when the training dataset trains the algorithm, the testing dataset can help in better accuracy judgment, as error can be easily identified and propagated. The training data needs to be carefully chosen because an imbalanced training dataset can cause imperfect decision boundaries by straining them due to lack of proper examples. Hence, the training dataset should be similar to the final dataset in its characteristics and should provide the algorithm with the labeled parameters required. An imbalance in this scenario would mean much more records for a particular type and very less records for another type.

The algorithm then finds connections between the parameters provided and essentially establishes a cause and effect relationship between the variables in the dataset. At the end of the training, the algorithm develops a draft of how the data works and the relationship between the input and the output.

Another important factor is to identify important features in the training data because irrelevant features can possibly influence the training of the algorithm. A few examples of supervised learning are regression, classification, and more.

Unsupervised Learning

Having a labeled dataset may not be possible in many cases. There is a cost of getting labeled data as well, as humans need to label the data. In the problem statements, where such constraints cannot be met, Unsupervised Machine Learning is the solution.

Unsupervised machine learning holds many advantages by being able to work with unlabeled data. Firstly, human effort is not required to make the dataset machine-readable that allows much larger datasets to be used by the program. Also, with the unlabeled data the algorithm has a better scope to draw conclusions and relationships in the feature set available. Another huge advantage can be its usage in real-time data, where input is analyzed and labeled with the learners.

The labels generated in the unsupervised learning work differently from the labels in supervised learning. In supervised learning, the decision boundaries have to be intently drafted based on the labels, whereas in unsupervised learning, the labels are an evolving space that can get updated with the drift of decision boundaries at a given time based on the hidden correlation it deduces.

The creation of these hidden structures makes the unsupervised learning algorithms versatile. Instead of a defined and set statement, the unsupervised learning algorithms can adapt to the data by dynamically changing hidden structures. These offer more post-deployment development than the supervised learning algorithms.

Clustering is one of the famous unsupervised learning algorithms, to split the data according to similarity, clustering can also help in identifying the anomalies in our dataset.

Reinforcement Learning

Reinforcement Learning is another Machine Learning domain that solves problem statements that are more of a path traversal rather than definitive output. Games and robotic movements can be a great example of using Reinforcement Learning. Tracing a set of paths such that system would provide reward for every correct path and penalize for every wrong path.

Reinforcement learning is all about learning and inspiring from experience and interaction. It features an algorithm that improves upon itself and learns from new situations like trial and error methods.

The rules of the game are not suggestive of the steps but rather a system learning of what kind of actions are a reward and what is the penalty. Such rules that are called Policy are often the most important part of such algorithms, where an agent is interacting with an environment. At each time step, the agent receives the environment’s current state, and the agent must choose an appropriate action in response. After the agent executes the action, the agent receives a reward and a new state. Basically, each iteration of the agent is analyzed by the interpreter to judge whether a penalty or reward is called for.

Figure 1.1: Reinforcement Learning

Applications of Machine Learning

Adaptability, versatility, and effectivity are the key features to sell any Machine Learning algorithm. The dynamic and high-powered nature of the versatile Machine Learning solutions is a feature utilized in situations where the answer is required to continue improving post-deployment.

Machine Learning algorithms and solutions are versatile and can be used to support and scale business teams. Be it image recognition, weather predictor, spam email filtering, fraud detection, and so on, Machine Learning has a huge scope in solving such problems. Cognitive assistants or virtual chatbots are another such example. Making any customer interaction a personal experience can significantly improve a business, so these chatbots analyze customer queries, preferences, and short context to curate an instant preset answer to assist the customer.

The customer profiling is another important application of Machine Learning. Streaming services like Netflix or Hulu curate a “You may also like this,” where they push suggestions based on the profile they have analyzed per customer. Finding items relative to each customer can be based on n number of parameters. The streaming services often consider the likes and interests of customers, whereas ecommerce services often curate “You may also like this” mostly based on the common preferences of similar customer profiles or rather the customers with similar wish lists or shopping patterns. Social Media on the other hand, curate advertisements, and news-feeds based on tentative likeability and revisits to certain domains of information.

Natural Language Processing

Natural Language Processing (NLP) is an important part of Computer Science in Artificial Intelligence (AI) that deals with how computers analyze, understand, and derive meaning from human language in a smart and practical way. Thus, enabling computers to understand and process human language.

With the help of NLP, we can build models and processes to analyze text. NLP has its application in numerous real-world scenarios like speech to text, language translation, text generation, summarization, sentiment analysis, topic modeling, text mining, automatic question answering, and so on.

Any language data, if deconstructed, is a mere stored stream of Unicode characters or entities for a system that consumes some physical space in bits and bytes all because machines do not understand any human language at the root level. NLP to a machine is bringing out sense in some stream of characters, purely by analyzing its constraints, rules, and patterns against some training data. The training helps in understanding of language syntax and semantics and generating complying information similar to that of the data that is trained.

Example: The word “ACT” has no isolated meaning by itself to a machine; it is not a keyword and clearly only consuming some memory. By analyzing English language documents and rules for its form and relative meaning to other words, we get that this English word “ACT” can have numerous word forms with several meanings in multiple contexts.

Each and every human language entity like the preceding example is holding a chain of information that makes NLP capable of many useful jobs like correcting grammar, converting speech to text, and automatically translating between languages.

NLP algorithms

Human languages are pretty complicated that make NLP difficult to learn and implement. In spite of that, NLP offers some techniques such as Word/Sentence Embeddings, Syntactic and semantic analysis, and so on that enables organizations to develop software that understands human language. Some of the heavily used NLP algorithms/steps are as follows:

• Pre-processing: The text that is received as human language often requires some sanitation, and there are several steps that can be incorporated.
  • Stop-words Removal: Removes redundant characters and words of the text that hold little meaningful information.
  • Stemming and Lemmatization: Converts words into their root form.
  • Tokenizer: Extracts token or phrases out of text (document or sentence).
• Parts of Speech Tagging: Parts of Speech (POS) tagging is an NLP technique that identifies or extracts structure of a sentence. For semantic analysis, POS tags or Parts of Speech Tags can determine what entities or tokens hold importance across the text. As an example, for a sentence “There is a seminar on Neurons and Neural Networks in Sydney,” the POS tagging for each of extracted tokens are as follows:
  “There(/EX) is(/VBZ) a(/DT) seminar(/NN) on(/IN) Neurons(/NNS) and(/CC) Neural(/NNP) networks(/NNS) in(/IN) Sydney(/NNP)”
• Entity Extraction: Named Entity Recognition (NER) is one of the most useful techniques in NLP that helps in extracting the entities from the text. It is generally based on grammar rules and supervised models. It highlights the essential concepts and references within the text. The NER labels classify atomic elements and important phrases within the sentence into categories like people, locations, organizations, dates, and more. Many cloud vendors such as IBM, Amazon, Google, Microsoft, and more provide NER feature as a service such that same can be leveraged in any application or product without a need for development from scratch. The Entity Recognition feature provided by these vendors have been trained on general English text or document such as Wikipedia articles.. Therefore, these models work best for identification of generic entities such as Person, Location etc. However, Google has released AutoML Natural Language service that enables developers or data scientists to train custom entit...