The human brain is one of the most sophisticated machines in the universe. It has evolved for thousands of years to its current state. As a result of continuous evolution, we are able to make sense of nature's inherent processes and understand cause and effect relationships. Based on this understanding, we are able to learn from nature and devise similar machines and mechanisms to constantly evolve and improve our lives. For example, the video cameras we use derived from the understanding of the human eye.Â
Fundamentally, human intelligence works on the paradigm of sense, store, process, and act. Through the sensory organs, we gather information about our surroundings, store the information (memory), process the information to form our beliefs/patterns/links, and use the information to act based on the situational context and stimulus.Â
Currently, we are at a very interesting juncture of evolution where the human race has found a way to store information in an electronic format. We are also trying to devise machines that imitate the human brain to be able to sense, store, and process information to make meaningful decisions and complement human abilities.Â
This introductory chapter will set the context for the convergence of human intelligence and machine intelligence at the onset of a data revolution. We have the ability to consume and process volumes of data that were never possible before. We will understand how our quality of life is the result of our decisive power and actions and how it translates to the machine world. We will understand the paradigm of Big Data along with its core attributes before diving into artificial intelligence (AI) and its basic fundamentals. We will conceptualize the Big Data frameworks and how those can be leveraged for building intelligence into machines. The chapter will end with some of the exciting applications of Big Data and AI.
We will cover the following topics in the chapter:
- Results pyramid
- Comparing the human and the electronic brain
- Overview of Big Data
The quality of human life is a factor of all the decisions we make. According to Partners in Leadership, the results we get (positive, negative, good, or bad) are a result of our actions, our actions are a result of the beliefs we hold, and the beliefs we hold are a result of our experiences. This is represented as a results pyramid as follows:
At the core of the results pyramid theory is the fact that it is certain that we cannot achieve better or different results with the same actions. Take an example of an organization that is unable to meets its goals and has diverted from its vision for a few quarters. This is a result of certain actions that the management and employees are taking. If the team continues to have same beliefs, which translate to similar actions, the company cannot see noticeable changes in its outcomes. In order to achieve the set goals, there needs to be a fundamental change in day-to-day actions for the team, which is only possible with a new set of beliefs. This means a cultural overhaul for the organization.
Similarly, at the core of computing evolution, man-made machines cannot evolve to be more effective and useful with the same outcomes (actions), models (beliefs), and data (experiences) that we have access to traditionally. We can evolve for the better if human intelligence and machine power start complementing each other.
While the machines are catching up fast in the quest for intelligence, nothing can come close to some of the capabilities that the human brain has.
The human brain has an incredible capability to gather sensory input using all the senses in parallel. We can see, hear, touch, taste, and smell at the same time, and process the input in real time. In terms of computer terminology, these are various data sources that stream information, and the brain has the capacity to process the data and convert it into information and knowledge. There is a level of sophistication and intelligence within the human brain to generate different responses to this input based on the situational context.
For example, if the outside temperature is very high and it is sensed by the skin, the brain generates triggers within the lymphatic system to generate sweat and bring the body temperature under control. Many of these responses are triggered in real time and without the need for conscious action.
The information collected from the sensory organs is stored consciously and subconsciously. The brain is very efficient at filtering out the information that is non-critical for survival. Although there is no confirmed value of the storage capacity in the human brain, it is believed that the storage capacity is similar to terabytes in computers. The brain's information retrieval mechanism is also highly sophisticated and efficient. The brain can retrieve relevant and related information based on context. It is understood that the brain stores information in the form of linked lists, where the objects are linked to each other by a relationship, which is one of the reasons for the availability of data as information and knowledge, to be usedÂ as and when required.
The human brain can read sensory input, use previously stored information, and make decisions within a fraction of a millisecond. This is possible due to a network of neurons and their interconnections. The human brain possesses about 100 billion neurons with one quadrillion connections known as synapses wiring these cells together. It coordinates hundreds of thousands of the body's internal and external processes in response to contextual information.
The human brain requires far less energy for sensing, storing, and processing information. The power requirement in calories (or watts) is insignificant compared to the equivalent power requirements for electronic machines. With growing amounts of data, along with the increasing requirement of processing power for artificial machines, we need to consider modeling energy utilization on the human brain. The computational model needs to fundamentally change towards quantum computing and eventually to bio-computing.
As the processing power increases with computers, the electronic brainâor computersâare much better when compared to the human brain in some aspects, as we will explore in the following sections.
The electronic brain (computers) can read and store high volumes of information at enormous speeds. Storage capacity is exponentially increasing. The information is easily replicated and transmitted from one place to another. The more information we have at our disposal for analysis, pattern, and model formation, the more accurate our predictions will be, and the machines will be much more intelligent. Information storage speed is consistent across machines when all factors are constant. However, in the case of the human brain, storage and processing capacities vary based on individuals.
The electronic brain can process information using brute force. A distributed computing system can scan/sort/calculate and run various types of compute on very large volumes of data within milliseconds. The human brain cannot match the brute force of computers.Â
Computers are very easy to network and collaborateÂ with in order to increase collective storage and processing power. The collective storage can collaborate in real time to produce intended outcomes. While human brains can collaborate, they cannot match the electronic brain in this aspect.Â
AI is finding and taking advantage of the best of both worlds in order to augment human capabilities. The sophistication and efficiency of the human brain and the brute force of computers combined together can result in intelligent machines that can solve some of the most challenging problems faced by human beings. At that point, the AI will complement human capabilities and will be a step closer to social inclusion and equanimity by facilitating collective intelligence. Examples include epidemic predictions, disease prevention based on DNA sampling and analysis, self driving cars, robots that work in hazardous conditions, and machine assistants for differently able people.
Taking a statistical and algorithmic approach to data in machine learning and AI has been popular for quite some time now. However, the capabilities and use cases were limited until the availability of large volumes of data along with massive processing speeds, which is called Big Data. We will understand some of the Big Data basics in the next section. The availability of Big Data has accelerated the growth and evolution of AI and machine learning applications. Here is a quick comparison of AI before and with with Big Data:
The primary goal of AI is to implement human-like intelligence in machines and to create systems that gather data, process it to create models (hypothesis), predict or influence outcomes, and ultimately improve human life. With Big Data at the core of the pyramid, we have the availability of massive datasets from heterogeneousÂ sources in real time. This promises to be a great foundation for an AI that really augments human existence:
"We don't have better algorithms, We just have more data."
- Peter Norvig, Research Director, Google
Data in dictionary terms is defined as facts and statistics collected together for reference or analysis. Storage mechanisms have greatly evolved with human evolutionâsculptures, handwritten texts on leaves, punch cards, magnetic tapes, hard drives, floppy disks, CDs, DVDs, SSDs, human DNA, and more. With each new medium, we are able to store more and more data in less space; it's a transition in the right direction. With the advent of the internet and the Internet of Things (IoT), data volumes have been growing exponentially.
Data volumes are exploding; more data has been created in the past two years than in the entire history of the human race.
The term Big Data was coined to represent growing volumes of data. Along with volume, the term also incorporates three more attributes, velocity, variety, and value, as follows:
- Volume: This represents the ever increasing and exponentially growing amount of data. We are now collecting data through more and more interfaces between man-made and natural objects. For example, a patient's routine visit to a clinic now generates electronic data in the tune of megabytes. An average smartphone user generates a data footprint of at least a few GB per day. A flight traveling from one point to another generates half a terabyte of data.
- Velocity: This represents the amount of data generated with respect to time and a need to analyze that data in near-real time for some mission critical operations. There are sensors that collect data from natural phenomenon, and the data is then processed to predict hurricanes/earthquakes. Healthcare is a great example of the velocity of the data generation; analysis and action is mission critical:
- Variety: This represents variety in data formats. Historically, most electronic datasets were structured and fit into database tables (columns and rows). However, more than 80% of the electronic data we now generate is not in structured format, for example, images, video files, and voice data files. With Big Data, we are in a position to analyze the vast majority of structured/unstructured and semi-structured datasets.Â
- Value: This is the most important aspect of Big Data. The data is only as valuable as its utilization in the generation of actionable insight. Remember the results pyramid where actions lead to results. There is no disagreement that data holds the key to actionable insight; however, systems need to evolve quickly to be able to analyze the data, understand the patterns within the data, and, based on the contextual details, provide solutions that ultimately create value.
The machines and mechanisms that store and process these huge amounts of data have evolved greatly over a period of time. Let us briefly look at the evolution of machines (for simplicity's sake, computers). For a major portion of their evolution, computers were dumb machines instead of intelligent machines. The basic building blocks of a computer are the CPU (Central Processing Unit), the RAM (temporary memory), and the disk (persistent storage). One of the core components of a CPU is an ALU (Arithmetic and Logic Unit). This is the component that is capable of performing the basic steps of mathematical calculations along with logical operations. With these basic capabilities in place, traditional computers evolved with greater and higher processing power. However, they were still dumb machines without any inherent intelligence. These computers were extremely good at following predefined instructions by using brute force and throwing errors or exceptions for scenarios that were not predefined. These computer programs could only answer specific questions they were meant to solve.
Although these machines could process lots of data and perform computationally heavy jobs, they would be always limited to what they were programmed to do. This is extremely limiting if we take the example of a self driving car. With a computer program working on predefined instructions, it would be nearly impossible to program the car to handle all situations, and the programming would take forever if we wanted to drive the car on ALL roads and in all situations.
This limitation of traditional computers to respond to unknown or non-programmed situations leads to the question: Can a machine be developed to think and evolve as humans do? Remember, when we learn to drive a car, we just drive it in a small amount of situations and on certain roads. Our brain is very quick to learn to react to new situations and trigger various actions (apply breaks, turn, accelerate, and so on). This curiosity resulted in the evolution of traditional computers into artificially intelligent machines.
Traditionally, AI systems have evolved based on the goal of creating expert systemsÂ that demonstrate intelligent behavior and learn with every interaction and outcome, similar to the human brain.Â
In the year 1956, the term artificial intelligence was coined. Although there were gradual steps and milestones on the way, the last decade of the 20th century marked remarkable advancements in AI techniques. In 1990, there were significant demonstrations of machine learning algorithms supported by case-based reasoning and natural language understanding and translations. Machine intelligence reached a major milestone when then World Chess Champion, Gary Kasparov, was beaten by Deep Blue in 1997. Ever since that remarkable feat, AI systems have greatly evolved to the extent that some experts have predicted that AIÂ will beat humans at everythingÂ eventually. In this book, we are going to look at the specifics of building intelligent systems and also understand the core techniques and available technologies. Together, we are going to be part of one of the greatest revolutions in human history.
Fundamentally, intelligence in general, and human intelligence in particular, is a constantly evolving phenomenon. It evolves through four Ps when applied to sensory input or data assets: Perceive, Process, Persist, and Perform. In order to develop artificial intelligence, we need to also model our machines with the same cyclical approach:
Here are some of the broad categories of human intelligence:
- Linguistic intelligence: Ability to associate words to objects and use language (vocabulary and grammar) to express meaning
- Logical intelligence: Ability to calculate, quantify, and perform mathematical operations and use basic and complex logic for inference
- Interpersonal and emotional intelligence: Ability to interact with other human beings and understand feelings and emotions
- Basic tasks:
- Common sense
- Natural language processing
- Intermediate tasks:
- Expert tasks:
- Financial analysis
- Scientific analysis
- Medical analysis
The fundamental difference between human intelligence and machine intelligence is the handling of basic and expert tasks. For human intelligence, basic tasks are easy to master and they are hardwired at birth. However, for machine intelligence, perception, reasoning, and natural language processing are some of the most computationally challenging and complex tasks.
In order to derive value from data that is high in volume, varies in its form and structure, and is generated with ever increasing velocity, there are two primary categories of framework that have emerged over a period of time. These are based on the consideration of the differential time at which the event occurs (data origin) and the time at which the data is available for analysis and action.
Traditionally, the data processing pipeline within data warehousing systems consisted of Extracting, Transforming, and Loading the data for analysis and actions (ETL). With the new paradigm of file-based distributed computing, there has been a shift in the ETL process sequence. Now the data is Extracted, Loaded, and Transformed repetitively for analysis (ELTTT) a number of times:
In batch processing, the data is collected from various sources in the staging areas and loaded and transformed with defined frequencies and schedules. In most use cases with batch processing, there is no critical need to process the data in real time or in near real time. As an example, the monthly report on a student's attendance data will be generated by a process (batch) at the end of a calendar month. This process will extract the data from source systems, load it, and transform it for various views and reports. One of the most popular batch processing frameworks is Apache Hadoop. It is a highly scalable, distributed/parallel processing framework. The primary building block of Hadoop is theÂ Hadoop Distributed File System.
As the name suggests, this is a wrapper filesystem which stores the data (structured/unstructured/semi-structured) in a distributed manner on data nodes within Hadoop. The processing that is applied on the data (instead of the data that is processed) is sent to the data on various nodes. Once the compute is performed by an individual node, the results are consolidated by the master process. In this paradigm of data-compute localization, Hadoop relies heavily on intermediate I/O operations on hard drive disks. As a result, extremely large volumes of data can be processed by Hadoop in a reliable manner at the cost of processing time. This framework is very suitable for extracting value from Big Data in batch mode.
While batch processing frameworks are good for most data warehousing use cases, there is a critical need for processing the data and generating actionable insight as soon as the data is available. For example, in a credit card fraud detection system, the alert should be generated as soon as the first instance of logged malicious activity. There is no value if the actionable insight (denying the transaction) is available as a result of the end-of-month batch process. The idea of a real-time processing framework is to reduce latency between event time and processing time. In an ideal system, the expectation would be zero differential between the event time and the processing time. However, the time difference is a function of the data source input, execution engine, network bandwidth, and hardware. Real-time processing frameworks achieve low latency with minimal I/O by relying on in-memory computing in a distributed manner. Some of the most popular real-time processing frameworks are:
- Apache Spark: This is a distributed execution engine that relies on in-memory processing based on fault tolerant data abstractions named RDDs (ResilientÂ Distributed Datasets).
- Apache Storm: This is a framework for distributed real-time computation. Storm applications are designed to easily process unbounded streams, which generate event data at a very high velocity.
- Apache Flink: This is a framework for efficient, distributed, high volume data processing. The key feature of Flink is automatic program optimization. Flink provides native support for massively iterative, compute intensive algorithms.
As the ecosystem is evolving, there are many more frameworks available for batch and real-time processing. Going back to the machine intelligence evolution cycle (Perceive, Process, Persist, Perform), we are going to leverage these frameworks to create programs that work on Big Data, take an algorithmic approach to filter relevant data, generate models based on the patterns within the data, and derive actionable insight and predictions that ultimately lead to value from the data assets.
At this juncture of technological evolution, where we have the availability ofÂ systems that gather large volumes of data from heterogeneous sources, along withÂ systems that store these large volumes of data at ever reducing costs, we can derive value in the form of insight into the data and build intelligent machines that can trigger actions resulting in the betterment of human life. We need to use an algorithmic approach with the massive data and compute assets we have at our disposal. Leveraging a combination of human intelligence, large volumes of data, and distributed computing power, we can create expert systems which can be used as an advantage to lead the human race to a better future.
While we are in the infancy of developments in AI, here are some of the basic areas in which significant research and breakthroughs are happening:
- Natural language processing: Facilitates interactions between computers and human languages.Â
- Fuzzy logic systems: These are based on the degrees of truth instead of programming for all situations with IF/ELSE logic. These systems can control machines and consumer products based on acceptable reasoning.Â
- Intelligent robotics: These are mechanical devices that can perform mundane or hazardous repetitive tasks.Â
- Expert systems: These are systems or applications that solve complex problems in a specific domain. They are capable of advising, diagnosing, and predicting results based on the knowledge base and models.Â
Here is a small recap of what we covered in the chapter:
Q: What is a results pyramid?
A: The results we get (man or machine) are an outcome of our experiences (data), beliefs (models), and actions. If we need to change the results, we need different (better) sets of data, models, and actions.Â
Q: How is this paradigm applicable to AI and Big Data?
A: In order to improve our lives, we need intelligent systems. With the advent of Big Data, there has been a boost to the theory of machine learning and AI due to the availability of huge volumes of data and increasing processing power. We are on the verge of getting better results for humanity as a result of the convergence of machine intelligence and Big Data.Â
Q: What are the basic categories of Big Data frameworks?
A: Based on the differentials between the event time and processing time, there are two types of framework: batch processing and real-time processing.Â
Q: What is the goal of AI?
A:Â The fundamental goal of AI is to augment and complement human life.
Q: What is the difference between machine learning and AI?
A:Â Machine learning is a core concept which is integral to AI. In machine learning, the conceptual models are trained based on data and the models can predict outcomes for the new datasets. AI systems try to emulate human cognitive abilities and are context sensitive. Depending on the context, AI systems can change their behaviors and outcomes to best suit the decisions and actions the human brain would take.
Have a look at the following diagram for a better understanding:
In this chapter, we understood the concept of the results pyramid, which is a model for the continuous improvement of human life and striving to get better results with an improved understanding of the world based on data (experiences), which shape our models (beliefs). With the convergence of the evolving human brain and computers, we know that the best of both worlds can really improve our lives. We have seen how computers have evolved from dumb to intelligent machines and we provided a high-level overview of intelligence and Big Data, along with types of processing frameworks.
With this introduction and context, in subsequent chapters in this book, we are going to take a deep dive into the core concepts of taking an algorithmic approach to data and the basics of machine learning with illustrative algorithms. We will implement these algorithms with available frameworks and illustrate this with code samples.