a INTRODUCTION TO = ARTIFICIAL INTELLIGENCE Oy ) Aye fp] To understand and appreciate Artificial Intelligence and describe its applications in daily ite io To understand the process of decision-making ed Introduction to Al-related terminology such as Al, ML, DL and domains—Data, CV, NLP {| Basias of Al Ethics BACKGROUND AND ContTEXT “The future is already here—it's just not very evenly distributed.” —William Gibson FOUNDATIONAL CONCEPTS OF AI We have learnt in the previous classes that Artificial Intelligence (Al) is not a new technology and has been around for a long time. The application of Al is also quite visible around us — in voice-enabled assistants, recommendation systems and in auto-complete suggestions when we type in Google. Googlehe basic concepts around intel}, ry and learn how we can Sree us in decision-making, That hich can independeng, il, some of tl .d. We will also t s which help It now becomes important for us to learn and how they are defined and understoo' human brain thinking and learning processe guide us in developing artificial version of intelligence or Al wi decision-making just like humans. Imagine having 2” Al system which can think o, “0 xplicit inputs or coding done its y .s without any Ms, own, learn on its own and take decision Sounds interesting? Let’s get started. words — Artificial and Intelligence. Artificial ig @ a nor natural, but Intelligence is a slightly com y Quotient) while others think of it ae nat wisdom oF knowledge ig es information and knowledge eal Artificial Intelligence comprises two to understand as what is not huma! term. Some people relate it to 1Q (Intelligence ability to remember information. Some also believe t intelligence. Can you explain thi e difference between ir first used in the early liam Stern. A French psychologist, Alfred Binet, igence tests to help the French governm fy children who needed extra nce. It was then introduced as the academic support and guidat concept of mental age or a set of abilities that children of a Alfred Binet and Wiliam Sten certain age possess. nce Quotient or |Q was The term Intellige gist Wil Let us first understand what is really meant by intelligence. Defining Intelligence people have understood intelligence in different ways. While ly refer to the following abilities: iring knowledge and the ability to remember it anc Over the course of time, definitions may vary, they broad! 1. Learning from experience: Acqu use it for future reference. 2. Recognizing issues and problems: Ability to foresee problems and ide! an early stage before they become too big to handle. Ey Problem-solving: Using knowledge and experience to solve a problem efficient manner. ntify them at in the most "The measure of intelligence is the ability to change.’hligence means presence of the above abilities along with skills around reasoning, te . ‘ fe ic, planning, etc. It has remained one of the most researched areas in the world and ic, : : pe interesting domain for computing as well. is try searching “Intelligence” on Google. It yields 1,04,300 crore results! according to researchers, intelligence is the “ability to perceive or infer information, and to retain it as knowledge to be applied towards adaptive behaviours within an environment or context.” — How 7 7 Tis ability to The ability Psychologists recognize problems 46 lam © ie define The abitty to Intelligence Solve problems We have seen that intelligence refers to a set of various abilities and skills. Each person may have different levels of these skills and abilities. You would have seen this around yourself as well — you might be good at drawing while one of your friends may be good at mathematics or someone may be a good singer. Sounds familiar? In a nutshell, we may define intelligence as: & Ability to Interact with the Real World: It is the ability to take action. For example, Speech Recognition, Image Recognition, cause and effect, etc. & Reasoning and Logic: * Modelling output, given input(s) ¢ Identifying and solving new problems * Planning and making decisions * Ability to deal with unexpected problems, uncertainties@ Learning and Adapre te eee . ae Adaptive Learning — our ee ex eriencing. Th S Ubg, for example; 4 child learns by observing a is i, ig. They see a 4, Beg next time can identify 4 treefleavesiflowers on cul oy Now, upon ~ a4 new variety of tree or flower, the learning model is updated internally cing any help. “tho e defined as: pes of intelligence ical Reasonin id mathematical logics ' they There are nine ty! 1g: Ability to understan\ mathematical Logi rmulae and expressions Linguistic Intelligence: Language skills — reading, writing, listening ang A ; oh telligence: Perceiving the visual world around us and the re, aking ation, Shi nt objects and things s of the position and movement of b 0 oY ban s 1. fo 3. spatial Visual Ini between differe _ Kinaesthetic Intelligence: by means of sensory organs Ability to create, identify and understand musical note S, thy Vn Awarenes: 5, Musical Intelligence: and patterns Ability of self-awareness and knowing about One’s On 6. Intrapersonal Intelligence: self — skills, abilities, weaknesses, Awareness of spiritual and ontology-related are; as 7. Existential Intelligence: 8. Naturalist Intelligence: Processing information around the environment a 9. Interpersonal Intelligence: nd Nature they feel and perceive s us understand that if we are to create a true Al system (the one th ne that pas, ese components of intelligence ne Not etc. Communicating with others with an understandin 18 Of hoy This help: the Turing Test, remember?), it should have all th just a simple high IQ or good processing power and numerical abili ility. The inte Faction with such an Al system should mi ake one feel as clo i bi ; A a : se as possible , Sa abilities and emotional capabilities and paremied ™ eae m should be our aim with the considerati eaitaled erations of ethics and bi oe ias to be equitd: and fair to all humanity. So, what is Artificial Intelligence? We have read in ; Conipiier eat classes that Artificial Intelligence is a field to do certain tasks aed developing machines which are i a : vee only by humans. It uses oul normally be considered w eon a machine learning to a teeieeae : b a arn and adapt!" algorithms t " 0 become without h smarter each luman intervention oan and autonomously un ily.- w Machines Become Artificially Intelligent Ho’ ou remember how we learnt to do things while we were babies? Initially a baby 0 gles to walk. He takes help from others while learning how to walk and once he snows it, he keeps on upgrading it by learning how to run, jump, etc. similarly, machines also become intelligent once they are trained with some information su , which helps them achieve their tasks. Al machines also keep updating their knowledge ptimize their output. to OF i eo EO Co | + Good for simple * Can handle complex = Able to understand * Human-level intelligence Good |" Gasetication and pattern classification tasks human motives and that can bypass our | recognition tasks * Able to use historical reasoning. Gan deliver intelligence too « Great for scenarios where data to make predictions personal experience to all parameters are KNOWN; Capable of complex tasks **VOH€ based on their | can beat humans Because such as self-driving cars, "eS and needs | itcan make calculations bur sti vulnerable to + Able to learn with fewer much faster outliers or adversarial examples because it | + Incapable of dealing examples understands motive and with scenarios including —. This ig the current state (Ment Imperfect information of Al, and some say we * Considered the next ‘or requiring historical have hit a wat! milestone for Al's understanding evolution Fig. 1.2: Types of Al Can you think of devices and products around your daily life which use Al? List them down and also explain the Al in them. For example, you may say that Alexa at home uses Al (Natural Language Processing) to understand and interpret our requests and play music or creates a personalized recommendation list of our favourite music playlist. How We Make Decisions Itis important to understand what is the use of intelligence. You might have heard the phrase “We are the choices we make” or simply put, your life is a series of events and how you respond to them and what “decisions” you take. A person who is more intelligent is likely to take better decisions and hence lead a better life for themselves and people around them. A great leader needs to be intelligent to make the right decisions for a corporation or a country. We are constantly making decisions. Stop for a minute and think how many decisions you take every minute. It is estimated that an average human makes 35,000 decisions a day. We make decisions every day from what to wear to what to watch and what to play, etc Have you ever wondered how do we make decisions or how our brain makes them for us? Our brain makes decisions based on past experiences and availability of information which includes self-awareness, intuition and knowledge. It also uses the nine intelligence skills an¢ abilities mentioned earlier in the chapter.be relevant and develop relationships and insights in orga, pe an z information must . of information is just as ble in our decision making, be.» the quality of NG i JS lust as impor Oe | valuable in our ce . king is “unknown’, we , ‘a he quantity. lf the key factor in decision making ; a © Might enco, Hf the quantit veperience ot information « Wty. * ons about which we have no prior experience or infe on and there mj rit situations 9 de “ 1 ituat! jsion and so on hth, of taking the wrong dec significant risk OR CR ON Cima of Artificial Intelligence is that people conclude iy MeO understand it” eae Let us understand this from these two scenarios: pau | MAKE YOUR CHOICES! | | Scenario 1 You are locked in the house and there are three possible exits but you can't open them on your own. The first | exit leads to. a jungle which has dangerous carnivores. | The second exit leads to a construction site with alot of | moving equipment and the third exit leads to a closed road - with no way to go. Which exit will you take? i Hint: Exit 1 leaves you with no possibility of help and is very risky. Exit 3 leads to nothing and you | can’t call anyone for help. What about Exit 29 There is a possibility of human contact and the | equipment can help in breaking the door to get out What do you think about this situation? | Scenario 2 Amit had a party at home with a few of his friends and when everyone left, he realized that his phone was missing. He began thinking | who could it be or did he by mistake lose it somewhere else. He looked at the CCTV footage and realized that he gave his phone to one of the guests to make a call and that guest left early and had the phone in his hand while leaving. The data helped Amit find the right person who took his phone As you can see from the above scenarios that information helps us to m | Do you agree with this approach? If not, what would be your appr LO od decisionsee ApPLICATIONS OF AI AROUND US Google Assistant its functions, Google Assistant relies on Artificial | | ree Google intelligence (Al) technologies such as Natural Language Processing and Machine Learning to understand what the user is saying, | and to make suggestions. Google Assistant can open apps, make suggestions, tell you the weather, play music, etc. Google Assistant supports 30 languages and ; has six different voices to choose from—one of which is that of award-winning musician john Legend. Google Maps Google Maps uses Al technology, known as global localization, to scan “tens of billions” of Street View images to help understand a person’s orientation | and the precise altitude and placement of an object inside a building. Google Maps: Recommendation System Arecommendation engine is an information filtering system rendering information tailored to users’ interests and preferences. It is able to predict a specific user’s preference for an item based on their profile. Amazon uses recommendations for targeted marketing via email campaigns and website pages. COLLABORATIVE FILTERING | CONTENT-BASED FILTERING ‘| Read by both users | = re) Read by user | Q=- | aid | | Recommended to user Chatbots oa | Chatbots are essentially software applications that use Al and NLP Helio! to assist humans and communicate through text or voice. They are virtual companions that seamlessly integrate into websites, | applications and even instant messengers, and help businesses get closer to their users. |hey can be categorized as: itten keywords that they understang Depending on how they are programmed, t ow must be coded into them 4, ah « Simple chatbots: They work on pre-wrl of the commands that they are going to follow 7 aet developer. So, if a user asks them something outside of their knowledge base the y respond with “sorry, | did not understand”, OF something along those lines + Smart chatbots: Smart chatbots are the F28° today! Based on Al, these bots have pre-programmed answers. They learn with time, catching keywords ang Pui them in context, and help users arrive at the most relevant answers to their queries uid consider. Apart from just getting rig ; ce, chatbots also go a long ways ore, users are finding it an amas ig that as these bots get me ‘e them with open ang s shot queries at on benefits and m Chatbots are something every busines: routine tasks and processing multiple gain customers’ loyalty. With all these , experience to get assisted by 4 chatbot. There’s no denyin ber of businesses will embrac sophisticated, a lager andlrgersue OS Play with Chatbots Find out chatbots on the internet and interact with them to understand and appreciate how they use Natural Language Processing (NLP) to have an engaging conversation with users, Sophia Sophia is a social humanoid robot developed by the Hong Kong-based company Hanson Robotics. Sophia’s architecture includes scripting software, a chat system and OpenCog, an Al system designed for general reasoning. Sophia uses speech recognition technology from Alphabet Inc. and is designed to get smarter over time. Its speech synthesis ability is provided by CereProc’s text-to-speech engine. ‘work traffic, * Cybersecurity professionals use analytics to detect exceptions in network patterns, n: Real-time and normal user activities. Hackers can be identified by their signatures or using Al can help identify any such signatures and detect an intrusion alert, | monitoring * Do you know that Sophia was the first bot that was recently offered Saudi Arabian cit izenship! Scan OR code or visit: https:/www.youtube.com/watch ?v=IsFv_gKS3YEa what is Not Artificial Intelligence we saw what is Al and how it is related to intelligence. We also remember that Turing Test helps us understand what truly is Al and what is not. Many a time, we come across advertisements or marketing promotions claiming that a certain product or device is atdriven. It is important to figure out if it is really Al or just an automated system based on rules known as Rule Engine. infosys CEO Vishal Sikka sums up Al as “ any activity that used to only be done via human intelligence that now can be executed by a computer,” including speech recognition, machine learning and Natural Language Processing. You would have heard about fully automatic washing machines or may have seen one being used at home. You need to adjust the settings of the machine and also make configurations for different sets of clothes and fabric type. It is an example of automation, not Al. You would have heard about fans and ACs that are Inte ‘ernet of Things (loT) enabled but you need to manage their settings for “On” and “otf” - You would have also heard about various robots for household tasks but they all need to be trained in some way at the start, right? Many times, there are smart systems around us which people feel are Al but they are not actually Al. For example, a smart refrigerator that adjusts the temperature based on outside temperature by a pre-defined logic. It is important to understand the devices and machines ar ‘ound us, be it our computer, smartphone or TV. We need to see if it js able to “think” * Nature of tasks changes + Nature of tasks changes + Decisions are automated =| + Humans inform machines * Machines learn continuously Assisted intelligence » Nature of tasks doesn’t change * Tasks are automated » Humans don’t learn * Machines inform humans » Examples: business strategy analysis using machine * Examples: autonomous vehicles, smart investment On its own or it is simply acting upon a set of instructions. | & ___ Man-machine Inteligence Continuum BQ | l EP | | Autonomous intelligence | Augmented intelligence | + Machines learn learning, smart clinical | ir ision su | = Examples: machinery, decision support | Processes in factories, boilers, | ovens Fig. 1.3: Evolution of Al Consider that a smart key is not really Al because it uses RF (Radio Frequency) sensors to unlock and open the car. meteAs you can see, Artificial Intelligence is a vast domain. Before we get into further oe | let us recap our existing knowledge of Al. ails, What do you know about Artificial Intelligence (Al)? What do you want to know about Al? What have you learnt about Al? Let us revisit the game of Rock Paper Scissors against an Al model. The challenge here is to win 20 games against Al before Al wins them against you, Scan QR code or visit: https://www. afiniti.com/corporate/rock-paper-scissors Did you manage to win or lose? What was the strategy that you applied to win this game against the Al machine? Was it different playing Rock Paper Scissors with an Al machine as compared to a human? What approach was the machine following while playing against you? List down your observations.S poRTANCE OF ARTIFICIAL INTELLIGENCE JM! coogle ¢ the other £0 Sundar Pichai said, “Al will have a bigger impact on humanity than some of well-known innovations.” ugy is one of the most important things humanity is working on. It is more profound than electricity or fire.” —Sundar Pichai al has managed to disrupt the conventional functioning of most of the work sectors and continues to expand its influence over the remaining ones. The most interesting aspect of Alis that it enables algorithms to handle situations and solve problems on their own without human supervision and learns on its own as well to improve over time using machine learning. Sounds amazing, isn’t it? Future will not be about Al vs Non-Al but people who know Al vs those who don’t and guess who will win! Summarize a list of daily activities you do which you believe can use Artificial Intelligence in future. | Think of a problem you face in your daily life—it could be related to travelling or at school or. / your home. Now, imagine a device that can solve this problem using Al. | EVOLUTION OF ARTIFICIAL INTELLIGENCE You might recall from our earlier discussions that Artificial Intelligence is by no neans a recently discovered technology but, in fact, started way back in the 1950s in he US and has been constantly evolving ever since. It has recently gained prominence ecause of internet penetration, proliferation of connected devices and data vailability. ‘om science fiction in the 1940s when Al first captured human imagination to 1956 artmouth Conference and then Turing Test to chatbots and IBM Watson now, Al has ken over six decades of research to reach where it is today.INDUSTRY 4.0 INDUSTRY 2.0 Om De INDUSTRY 1.0 Te Rem Sat ca Beh Sy Beem Ca) nbly line, os Sa Moe PN a d Cele eure Fig. 1.4: Industrial Revolution Over the Years (Source: ICT News) Why AI is Gaining So Much Relevance Today The key reasons for this recent popularity are: & Al systems are resource-heavy. We now have increased and more than ever powerfy| computational resources which enable the heavy Al algorithms to work. Do you know that today an average smartphone has more computational power than what was used on the entire Apollo mission which got man on the moon! Al and machine learning need lots of training data to become better with time. Earlier it was difficult to consolidate data into a single platform. Now we have so much data available with billions of people using cheaper smartphones and so many apps on them constantly generating more personal data. Internet and data plans have become cheaper (Remember Jio?) with more internet Penetration in the last two years making more and more people and devices to get connected and enable Al globally. We all have access to apps such as Cortana/Siri/Google Assistant that know so much about us—our choices, preferences, location, profile, etc.—to serve relevant content and information to us enabling shift from narrow Al to Broad Al. (Do you remember the difference?) Commercial data sets are available to support analytics. Do you know about Government of india data sets, Google Images data sets, etc.? Many of these are available as cloud services and APIs for ready integration. Machine learning algorithms have been made available through open-source communities with large user bases. There are more resources, frameworks and libraries that hav? made development easier. Have you heard about a framework called TensorFlow? Data Visualization is more user-friendly with tools such as Tableau and microsoft Power BI. It is easy to run data science models and interpret them in a rich use™ interactive mode. (We will read about it in Chapter 4.)* applcatiO gramming interface (API) is a computing interface for interactions between | - asia software layers/programs It is like a language for software to talk to each other. More | mcdanacatty, a comprises C2) requests, their methods, data formats and other syntax ) Artificial! Intelligence (Al) Trends Independent of Application - emehnasin e (2050) [ase W~ A Seig Wo oecite 2 ighe task beter than ] JZ Future (2050) “Smtr ) 4 7 General Al Acceleration * (with or without Possible | consciousness) Narrow Al Superclusters ‘ Convergences Driving Acceleration -_ * Internet of Things 7 Enhanced + Next-Gen Chips (Moore's Law 2.0) ¢ Narrow Al * Supercomputers * Sensors MILESTONES AND CAPABILITIES ‘\ 4 cd UberBots y (’Alfirst”) Robotics Cloud Narrow Al i Deep Learning (imagery, language. etc ) Bots (conversational, search, etc. 2017 201 2021 _ § ee FutureGrasp, LLC Fig. 1.5: Evolution of Artificial Intelligence Al can have two purposes. One is to use the power of computers to augment human thinking, just as we use motors to augment human or horse power. Hobotics and expert systems are major branches of that. The other is to ase 4 compater’s artificial intelligence to understand how humans think. te 2 humanoid way. If you test your programs not merely by what they can accomplish, but how they accomplish it, then you're really ving cognitive science; you're using Al to understand the if homan mind” ~Herbert A. Simon7 yin wield He abe 4p he Hipihet d You read this quote in class 9 as well Hut M th a alte, he rea ap the concept of decision malig, andl how butt anal ie UHI pr Purge ty developments is to really understand hunt inal | alan 3 I06 7 my we do that, we can simply create ary artifielal praeeee UF SY It to rey hy distribute and scale It Cate The Possibility of a True Al System Humans have always trled to make machines think Wore and more tke us ae 0 it lon My is fascinating in the way it works, Earlier, we only HMiougl! f machines to hep "bray yy vt Al, are Joking gr ig 7 Ot mae labour-intensive work or hard tasks but increasity to truly replicate our thinking process. In suc fa world, humans and hachines county ul collaborate with each other. ey, Ua ST Currently Eres Technology * Machine Intelligence is here Fig. 1.6: Stages of Al (Source: wany xenonstack com Before moving any further, let us do a quick recap of definitions of Artifici Intelligence (Al), Machine Learning (ML) and Deep Learning (DL): & Artificial Intelligence or Al refers to a set of techniques that enables s to mimi human intelligence and be autonomous. An artificially intelligent machine works . algorithms and data fed into it and gives the desired output out any guidance or supervision. Machine Learning or ML is a subset of Al that enables machines to improve with experience or more data sets. The machine learns from training data set and then improves with the test data using multiple iterations. It mainly focuses oncom uter algorithms that improve automatically through experience. : . Deep Learning or DL enables the machine to train itself to Perform tasks with iarge amounts of data. It is the most advanced field of study within Al a a aine at replicating human learning process. In deep learning, we ae coe Ge Go Program as machines are intelligent enough to develop algorithms ¢o’ s igorithms for themse!yelligence is the high-level umbrel rm which nit ell intellig i igh-level lla te hich holds both deep learning as w g. Deep learning, on the other hand, is a very specific learning approach f machine learning as it comprises multiple machine learning algorithms. nS ag aT Tey Subset of machine learning in which multilayered neural networks learn from vast amounts of data «ial rtifcial ; an jacine Fearn * risa subset © whic Maa INTELLIGENCE ‘program that can sense, = reason, act and adapt Algorithms whose performance improves as they are exposed to more data over time Fig. 1.7: Al, ML and OL pasics of AI - It all Starts with Algorithm itis important to understand that we need to start thinking about these abstract concepts in a more algorithmic manner. So, in subsequent chapters, we will be able to create functions around it and then a code. What is an Algorithm In Mathematics and Computer Science, an algorithm is a structured sequence of well-defined, computer-implementable instructions to solve problems or to perform a computation. We will be using lots of Modelling in Al which is essentially developing algorithms which can be trained to get intelligent outputs and we can define an output y=f(x) for any set of x inputs. In previous classes, we have seen various types of graphical representations and data visualization techniques which can be used for representing different sets of data. We will see that Data becomes the starting point in most Al systems and models. You must have heard the term Big Data, right? Fig. 1.8: Algorithm Type Big Data analytics is not a new technique. In fact, its earliest roots can be traced back to almost 18,000 years as a science of analyzing data beginning with scratchings on bone Sticks in Africa. The earliest example of Big Data in our current civilization can be thought of around 1880 census when IBM used business intelligence to assist the US government reducing the analysis time of the census from an expected 10 years to only three months.systems Design Is: y around the w orld in vario ae The simplest framework in Al a © Understand: There is? lot onesie ma etuctred a io, video, etc-)- i ’ : (doe Pee aferer ), interpreting the a is the “ a Bye, ®, Reason: Deriving patterns from the data set is a cura the process includes using various methods of classifying and visuallZ s : ©. Learn: Doing iterations and constantly updating the analytical model with new data to see if the model holds true or not. . Pattern Recognition of the Al devices around us is to understand and ana, t by using a process re ean outpu h supervised and unsupervised leap, d underlying patterns in ¢ pect of most ages of tern recogni mber the difference? videos and produc ition uses bot ) to fin One common as data, such as text or im pattern recognition. Pat techniques (Do you remel derive insights. Do you think rec List down what you think. in YouTube works on pattern recognition? ommendation system it ee around yourself are based on ies of pattern recognition you 5! hms. What it does is that it trains an Al model on a number of ase you want to train a model to recognize your tried it? It is fun.). You might have experienced m used by Google, Linkedin and The most common exampl supervised learning algoritl s. Let us say your face images in ¢ (Have you ended response syste! data set: face to unlock your smartphone an example of this in the recomm Gmail on communication. nique known as fuzzy logic where answers Most pattern recognition algorithms use a techi would be the closest match but never an exact match. If you asked an Al system to add “Well, based on other addition examples I’ve been given, it’s 1 +1, the answer might be close enough to the value of 2 to be 2. So, !’m saying 2.” Imagine, if you said this in your class to your teacher! © 9 as well, most examples of Al around us are of Narrow | Artificial General Intelligence As we read in Class ntelligencé : rtelligence. The ultimate goal of Al is towards achieving (acl) with ethics and explainability into it (we will read about these terms shortly). 4yman ine rter. tions AGl aims at having machines at least as smart as humans and maybe even sma brains, described in computer terms, can perform 38 thousand trillion opera second. Our fastest computers are still a little far away. a| know about the Open Mind Common Sense project at MIT that has collected more than @ ee 0 vor re it started in 1999. This particular database contains statements of facts that we might consider De These short sentences are later parsed into structured data so they can be easily understood ed by a computer in the same way it might search a library catalogue for a book author based on an the title of the book. facts si the final output of pattern recognition is inference. To understand it in a simple way, \et’s say your mother gives you hot food at lunch. Now, if the vegetable is hot and you yse a spoon to eat it, the spoon will also become hot (unless it is made of some special heat-resistant material). The only problem is that most of the information is unstructured (unlabelled, unexplained, uncategorized and in various non-usable formats). A lot of researchers are working on developing complex Al systems to understand and interpret such unstructured data with faster processing speed. the accuracy of such pattern recognition algorithms is constantly improving but we need to be careful of false positives and negatives (We will talk about it in Chapter 7). Another important aspect is the explainability. For example, a recent application of deep learning to cancer detection found the algorithm more accurate than a certified practitioner. But unlike a doctor who can give explanations of reasons leading to their diagnosis, Al system is not able to explain its rationale always. This is called black box problem or lack of explainability. at o Deep learning networks are so complicated that new methods of computation—GPUs—are required to build them. What is a GPU A Graphics Processing Unit (GPU) is a specialized rena ee unit designed to rapidly manipulate and ora manage memory to accelerate the creation of images in a frame buffer intended for output to a display device. They are used in embedded systems, workstations, mobile phones, personal computers and game consoles. GPUs are very efficient due to their highly parallel structure and are becoming more popular than general-purpose central processing units (CPUs) a for algorithms that process large blocks of data in parallel.i One of the core purposes of Al is to understand how ae ae wos te, Mimic, Create artificial brain-like intelligence. Cognitive Computing (ce eating Aor : My that can simulate the human brain thinking and learning Process: tworks (Remember we read about it in ee king process. It covers various subjects . erg ing (NLP), human-machine interface, & » Cognitive Computing Cognitive computing is based on neural net in Class 9) and deep learning to design thin Computer Vision (CV), Natural Language Processi puting is IBM’s Watson, which won the fa : game show Jeopardy! We engage with cognitive computing in our everyday lve with realizing it. When we are listening to a ong OF seeing a product On aN app, we st, seeing recommendations or similar things, right? These recommendations are Tender through sophisticated machine learning models and adapt to our behaviour. A well-known example of cognitive com| TYPES OF LEARNING We have talked about different types of lea Let’s quickly revisit them. rning used in machine learning algorithms Reasoning f achir Ta Natural Language Seeeesine NLP) [gp Supervised | Unsupervised JReinforcemen Learning Learning Learning Planning Supervised Learning Most examples of learning around us follow a simple structure—we see something, label it with some attribute, and remember it for future reference. Sounds familiar? Let s ay you are seeing a computer for the first time. You would try to see its different parts (monitor, keyboard, CPU, etc.) and remember them so that the next time when you see a similar device, you know it is a computer. Supervised learning follows this principle and trains the model on a large data set classify the data and find patterns in it. It tries to establish a y=f(x) for a given set of (y, x) so the next time you see an x, you know the y using the function! You may remember our activities in Class 9 around mystery animal and similar games where Al tries to predict something based on previously trained data. The data M4 be continuous wherein we use regression and if it is a finite set of values, it is know" as classification. a aWe all hear about weather news daily which uses weather forecasting. It looks at past historical gata and various parameters (humidity, wind pressure, etc.) and predicts weather for the day. Itis important to make sure that we have enough training data. If the model is fit for only a few patterns based on a smaller data set, we have a problem called overfitting. It means that the model is tuned for training data only but may not work with the test data or new unknown data. supervised training models are used in various applications such as fraud detection, pattern recognition, recommendation engines, credit analytics, etc. Do you remember we also studied about Decision Trees in Class 9? Decision trees can be used | to set up what we call Expert Systems that work on simple yes and no rules. For example, to | enter a room, you need to check if the door is locked or not. If it is locked, you need a key | and if you don’t have the key, you need to open it by force. At each stage, you have multiple | options, right? This is how a decision tree logic works. Can you create a decision tree for someone trying to enter a room? To summarize, supervised learning works on the simple logic of “Show lots of images (till n® time) of cats and dogs to the algorithm and tell it which one is dog and which one is cat so next time or (n+1)" time, when you show the image, it will be able to classify it as dog or cat.” at Pe Labelled Data: Labelled data is a data set which has been tagged with one or more labels, for example, in email spam, labelling a mail as spam based on certain attributes such as subject links, content, etc. Unlabelled Data: Unlabelled data is the data that comes with no such tagging or labelling, | Be for €xample, a comment or picture on social media added by someone without any tagging orunsupervised Learning K, inkedin, Instagram, etc., Benerate social media apPlice"0" t users and thelr activities. This data may be jn .™%y f data every By : ete, andma not be labelled or taggeq a the i and classification with SUpervisey, ma sters OF BrOUPINg of dat, Sy eat,” re ications such abi amount 0 of comments, it difficult to ana! unsupervised learning inc! (We will talk about Data features 0 actually move ters establishe uch as clu 5 Data Science in Chapter 4 ) ludes techniques § when we discus The next step is t towards 2 supervised learning approach once , some broad parame! d through clustering. It is ay useful when wep e lag of unlabelled data to actually group it before creating 4 relationship among data fae he For example, in COVID-19, there are @ lot of symptoms (mild fever, aches, throae Ur etc.) and still it is difficult to label it with an outcome, fora large number of a iste ideally be solved using unsupervised = - mir asymptomatic. So, such problems should ised algorithms include the Apriori and the k-means, w, ng 8 i Some common unsuperv! cover some of them in subsequent chapters. Reinforcement Learning elled data is available and the situation keeps on ch, Cran ging s of data. Take, for example, a self-driving car. Itis ye, SIS Very Think of a scenario where NO labé own situations and the car constantly inte Facts generating new types and form: complicated as we may encounter unkn' with the environment. ehavioural learning model and on the basi is of cement learning is based on the b' ‘A deep learning agent is any autonomo Reinfor he environment ( interaction of the agent with t or semi-autonomous Al-driven system that uses deep learning to perform and imp ad improve upon its tasks.). There is always a base/start state and an end state for the agent andin reinforcement learning, the agent interacts with the environment to reach the end st ach d state uccess but is penalized for failure. So, in a way, the 7 and improves. The Al system or agent learns through r. A sequence of successful choices and paths will result in the process beh P if The agent gets reward for s learns from the environment and erro “reinforced” to get to the end state. In sit in re it raini something n simplest terms, you can relate it to training a dog. If the dog gets so thi g gets s ing to eat jo this trial and a pat on the back every ti i d ry time he sits or dos eran ere ee es as asked by the trainer, he will d from unstruct id of ayers ures cate Se Deep learning is similar to a typical ! eae umb is ‘al neura’ there will be in the model. that more complex the problem, more hidden 9!®nd Deep Learning etworks 4! | i rns. Neural eure! ‘ re a set of algorithms that are designed to recognize ate setae etworks aFe 2 Or and classify unlabelled data according t0 17 ariti | n mo : i ns. elses and deep learning maps inputs to outputs and finds correlatio! ant to understand the context etwor! ner ample inputs if a tree loses its leaves in the n the same tree data quality is found to be good, it is import aver the G28 applied to the problem. For example, cea ummer, it is a sign that the tree is unhealthy. But whe! vie a the middle of a cold winter day, it is considered a normal occurrence. ; ext of the data, you are likely to misinterpret i standing the cont Therefor i ee eee attention is paid to correlation between data ces hat are the relationships between conditions? In the example of the health : wee there is a direct correlation between the seasons and the colour and amount Z ae on the tree. But you also have to be careful about correlations. ioses its leave: Simple Neural Network Deep Learning Neural Network @ Output Layer @ Input Layer @ Hidden Layer You might find a correlation that makes no sense because the context is wrong. There may seem to be a correlation between leaves falling off the tree and the number of coats being purchased online. While both events are happening because the weather is colder, there is no relationship between trees and coats. Applications of Deep Learning Deep learning algorithms use imaging and sensors data to interpret the unstructured data to make near real-time autonomous decisions. 1. Self-driving cars use deep learning to help the vehicle understand the environment around the car. It is heavily used in health care industry for diagnosis of diseases at an early stage. The most popular application of deep learning is virtual assistants ranging from Alexa to Siri to Google Assistant. . It is used in Visual Recognition and fraud detection. itis involved in the generation of new sets of handwritings along with translation of images and text.Af = | j _— a ye a Thy Customers eep an existing custo Love Thy in marketing: “qt costs 4 lot less to keep mer than i be In There is 4 saying, anew customer. rketing departments across organizations, you will fin rac mer lifetime value (Ow much revenue is generate’ a rs) and attrition oF churn (customers leaving). While analyzing ru >” you have access to 4 wide variety of data these days aroun Ong ling oF patterns, preferences and service support data. It is now possible i al Ns data to create @ perfect personalized plan for the user. | can you think of ways in which Al can help in solving the problem of customer atrtons What data will you use? + Which data is structured and + Will you use supervised of uns « What key metrics will ir Use what is unstructured? supervised or reinforcement learning? k benefits? | you define to trac e used for object detection and Classification, w resolution aspects of the Images We ition include the eyes, nose, ni ety, actually create a model for recognan’ i ing can bt redundant data, 10 ements needed for facial recogni With more and more data, you can photo or CCTV recording as well. of the police department and have access to this technology. You a rds and images, CCTV footage data of criminals and suspicious aie ussed how deep learnii We have disc! del to remove can also tune the mot The most important el any special attributes. someone from a group Imagine you are in charge have access to criminal reco What would you do to set up a mechanism to catch cri + What data will you look at? + Will you use supervised, unsupervised or reinforcement learning? | © Describe the algorithm steps you will take to reach the criminals in any situation iminals? Machine Learning Algorithms Overview rn ji a ss ne ieee algorithms are a set of standard instructions for a computing w to handle data. An algorithm can be ir eet used for a simpl s ee et imple task such as multiplication Wotton ne 10 eee recognition. Machine learning algorithms are eee ludes access to various code librari pot platforms such as GitHub and Stack Overflow. endear Machine learnin; , ig algorithms are diff. In usual programmi ifferent from other st iminy i A andard programming alg transformed. In fel algorithm is created and data . oe oe oe more data, an algori learning, the data itself cr passes through it an gorithm becomes better and eates the model. With mo More accurate. 22 ft ia a aea lot of machine learning algorithms and selecting the right one for 8 given i . ; ere 2 ig actually based on experience and constraints. It is important to have an idea por gifferent classes of machine learning algorithms so that when the time comes, ui - 2b0 ‘no which one to select. you come of these are: a esiaD Theorem Notation vs Bal : | ai supervised learning. in machine ris it is required to select the best res (2) given the data (b). ; dassification problem, our hypothesis (a) in A be the class to assign for a new data instance (D)- Bayes’ Theorem provides a way that we can calculate the probability of a hypothesis given our | #1 prior knowledge. It depends on the conditional probability. Bayes’ Theorem is stated as: p(alb) = P(bla) * P(a) / P(b) where + P(alb) is the probability of hypothesis a given the data b. This is called the posterior probability. + P(bja) is the probability of data b given that the hypothesis a was true. * Pa) is the probability of hypothesis a being true (regardless of the data). This is called the prior probability of a. * P(b) is the probability of the data regardless of the hypothesis. Clustering We know clustering is used in unsupervised st learning. The idea is that the data points with us similar attributes are grouped together in 2 cluster. It is very helpful when we have a Be large unlabelled data set—it is a good place to uF start. 'Decision Tree Decision Tree is also an intere to come to a decision given yarlous root from where the decision tree Sturt which s further divided into nodes with € @ possible outcome and brane hes as You can assign & to likelihood of each oute This technique can be used for both eh problems. dimensionality Reduction ¥ ° | a — Y @ a> x 7 ’ Lh there are a few parameters like ignition o! We need to filter this to ensure th relevant data points remain. This can be Instance-based Algorithm Instance-based algorithms or lazy learners u: rize the new data set or test data set and catego! training data. sting supervise options with the & rach node fet the possibl ome given the jassification iat algori done through dime After 25 Instances y, technique ed tearnit with the like a tree antire data set presenting itis Je results/optlons: options: and regression pimensionalily Reduction The number of inputs present in a set is called dimensionality, This data : is interesting technique to remove Onset or redundant data. This is particularly Usefy a steady stream of data is bein when ated through devices or machines, gener For example, if you have an loT-based senso, in your car to fetch you data like car ignition on, GPS coordinates, driving speed, et, nn which will constantly give you “On” status, ithm speed is optimized and only consistent nsionality reduction. se training data to compare with the new data points based on similarity to the After 100 Instances 2a Y These algorithms can be very useful in pattern recognition where consistent data set flows exist rather than disparate non-continuous data. They are prevalent in pharma industries:al ) arization te Avoid Over regu g we discussed how overfitting is a problem in machine jearning with limited dats set. Regularization is a | sechnique to modify models to avoid the Problem of overfitting- Regularization simplifies overly complex | models that are prone to be overfit. This is e ‘entially 3 form of regression that regulari estimates towards zero. Ss the coefficient | Rule-based Machine Learning Rule-based machine learning algorithms use relational rules to describe data. With different set of inputs, output is described as a relation wit x). Rule-based machine learning approaches include learning classifier systems, association rule learning, artificial immune systems, and any other method that relies on a set of rules, each covering contextual knowledge. Simple Machine Learning First Generation Second Generation ith the inputs as Deep Learning Third Generation Training a machine learning algorithm is usually in three steps: 1. Representation The algorithm interprets the inputs to transform the data into an output. A y=f(x) is created to establish the relationship within the data set. 2. Evaluation As the algorithm creates multiple models, we will need to evaluate and score the models based on which model produces the most accurate predictions. 3. Optimization After the algorithm creates and scores multiple models, we need to select the best performing algorithm. As we have more and more test data, the model evolves. The ost important part of the training process is to have enough data so you are in a Position to test your model. Often, the first pass at training provides mixed results. This means that you might either need to refine your model or provide more data. Understanding your data is critical to your success. If you create a model based on faulty data, your predictions will obviously be inaccurate. In addition, you need to think about what data should be included in your machine learning application.following: ntification) readers, smart Examples of structured data include the medical devices and Globay : co, Ost, i * Sensor data: RFID (Radio Frequency Ide! (Programmable Logic controllers) controllers, System (GPS) data. ; + Server and log data: User logs, Serve data logs, security logs, etc. rading data. nting systems and t ata such as pressure, temperature, humidity ) Rte y logs are gener licks, etc. * Financial data: Accou * Weather data: Forecasting 4: ated as the user py, OWSe * Website/E-commerce: User activit} website—scrolling data, heat maps, ¢ d data include the fo! notes, hard copy data, Jlowing: Examples of unstructure: etc., in any Organization, Th + Thing * Undocumented data: Emails, of data generated by your class! access social media platforms such as Facey 0k, * Social media: Whenever users erated as well as photos vide ‘, » Videos Linkedin, etc., logs, trails and click data is ge"! etc., are uploaded which generate a lot of data. ges, notes, calendar inputs, pictures, y * Mobile data: This includes text messa bile applications. and data entered into third-party mo more CCTVs are being install led everywhere and the « Surveillance Data: More and generate images and recording data. In 1985, a group of researchers at Carnegie Mellon University created a chess-playing computer| called ChipTest that was capable of searching through 50,000 moves per second. It led to/ the development of next generation chess-playing computer called Deep Thought after te fictional computer in Douglas Adam’s The Hitchhiker's Guide to the Galaxy. Was it smarter than Garry Kasparov? Deep Blue* won using sheer brute force. Current versions of Deep Blue can evaluate more than 100,000 chess moves 2 second, where Garry could evaluate about three moves | a second. Supercomputing or fast computing is measured in gigaflops. One gigaflop equals one billion math operations 5 a second. “Deep Bive was the next generation of Deep Thought| {ris claimed by experts that Deep Blue didn’t use any Al; | humans in areas requiring strategy and skill. There was no deep learning in Deep Blue. Every move | had to be calculated at the moment ; it proved that computer could win over DOMAINS OF AI al has three underlying domains: Data Science, Computer Vision (CV) and Natural Language Processing (NLP). We will now look into details of each of these and then interpret them together in action as well. Data Science You are aware that as a society, We are generating data at an unprecedented rate. These data sets can be numeric ( sales, insurance Premium, weather data, etc.), categorical (colour, gender, etc.), or even unstructured free text (comments, notes, feedback). Data collection is the process of identifying various sources of data (structured and unstructured) and collecting data and Preparing to label it. We need to make sure that data collected is in the correct format and aligned with the project requirements. The key to data collection is to structure the data in the right format so that one can run analytics and pattern recognition models on the basis of this data. Predictive and forecasting models are completely dependent on the quality of input data and its consistency. “Garbage In, Garbage Out (GIGO)” holds true in this case. So, focus on ensuring that you get the right data set, clean it and check for consistency and hygiene. CU THE DATA SCIENCE HIERARCHY OF NEEDS al, DEEP LEARNING NB TESTING, EXPERIMENTATION, SIMPLE ML ALGORITHMS Learn/Optimize ‘Aguregate/Label ANALYTICS, METRICS, SEGMENTS, AGGREGATES, FEATURES, TRAINING DATA Explore/Transtorm INSTRUMENTATION, LOGGING, SENSORS, EXTERNAL DATA, USER GENERATED CONTENT m = (Source: The Ai Hierarchy of Needs’ by Monica Rogati) : Fig. 1.9: Monica Rogati’s Data Science Hierarchy of Needs> Monica Rogati’s Data Science Hierarchy of Needs is @ pyramid oe £0 underse with having, basic hardware, Sensor. ang intelligence levels in an organization. It starts a devices in places to capture the data required for our Al model. We need to ensure on data, right formats and quality of data at this stage: ‘ Next is having the right storage systems which means having IT infrastructure (se,,,_ cloud storage, etc.) and systems such as ERP, MIS systems which can keep this da," right data in correct format (structyy, eq We then need to do data cleaning to ensure that data) is available to run any analytics or data science models on it. Next step is to run data visualization models, classification of data, data labelling defining some analytics metrics forthe identified set of 0912. ms to identify patterns and forecas The final stage is to apply machine learning algorit future trends on data. We need to do A/B testing to iterate on Se Buus 7 form of rich media such as pictures, video ang : We have a lot of data available now in the fe * voice files. In the past, we used to minimize the amount of this type of data that we capturey : because we couldn't do much with it as it required large computational abilities, not to mentio, | the heavy storage cost of media files “Some people call this artificial intelligence, sae ed will enhance us. So instead of artificial Pol ow any Pym rae O a ae / —Ginni Rometty * Application of Data Science Price Comparison Websites This is an example of simple meta search price comparison technique —=—— | | Sen ny 4 21500 | | | | i}° * Initial price Best price gl ° a Discount Promotional price price Fig. 1.11: Data Science uses ML algorithms to forecast future values Website Recommendations This is a sample framework for setting up a recommendation engine by using collection -> storage -> analysis technique. Front-End A / sooo eh t % Machine Learning Fig. 1.12: Data Science uses recommender systems to predict user preferences BIG DATA AND AI There’s a mutual reciprocal relationship between Big Data and Al. Al depends heavily on Big Data structures for success, while also helping organizations unlock the potential in their data warehouses in ways that were previously very difficult or impossible to achieve. eet eo)Historically, when it came to analyzing data, the only option was to use 9 quer, SQL (a long list of complicated queries). Now with Big Data, we have newer metho Pre processing and analyzing data p Al is the next logical step to SQL-based search Organizations a to combine the power of human intuition with machine intelligence to augment these technologies or create what is called Augmented f, Intelligence. More specifically, an Al system needs to learn « x from data, as well as from humans, in order to be able to M iY fulfil its function, (Source: Marvvie Unwersity) oe You can use the following data sets available on Google Research for your mn Al projects, BE ae Scan QR code or visit: a 3 ewe http;//research.google/tools/datasets, ea § COMPUTER VISION (CV) Computer Vision (CV) has been around for nearly 50 years and its development started around the same time as Artificial intelligence in the 1950s. But like All, it is only recently that CV has started to gain importance. Some CV applications include self. driving cars, facial recognition-based tracking systems with vision cameras and Amazon Go au AERO a aes hee ears ie re rer ues Ce) SAU Ue ee eae ee aT queue, With the Just Walk Out Shopping experience, simply use the Amazon Go ° LCR La ace Te ne CORN om eats hs (Co sey fers aT Te aig Bee eRe ear RE are The key underlying technology in these amazing use cases is Computer Vision. So what exactly is Computer Vision? In simple terms, CV is a domain of Al that lets machines see and extract meaning from pixels in an image and afterwards Predict sor about it. decisions In the context of biological systems, CV aims to Mirror how human vi interprets things we see. The intent is to create autonomous system surpass capabilities of human vision, : 30 sion works and s which can even Escennarc apJearning can work hand in hand with cv creating powerful systems such as searching od in Google, tagging of friends in social media, apps which can create an aged ima sion of your face, speech to-text translation with native ve detection systems, etc. language speaker, Intrusion @ Algoritims — @ summa Nov ora | Tracthbla First use of | ee Leamiig | Fob 2015 | Machine accuracy | Surpat humans 0% L__ I \ 20102011 20122013, 2014 2015, 2016 Fig. 1.13: Computer Vision vs Human Vision Error Comparison APPLICATIONS OF CV Web: From traditional text searches, we have moved onto ima; object recognition, scene recognition, geolocalization from visi media, Google maps aerial imaging, ge search, face recognition, ion, image tagging on social YouTube (content categorization) VA/AR: Outside-in tracking, inside-out tracking (simultaneous localization and mapping, HoloLens), object occlusion (dense depth estimation) Smartphones: QR codes, computational photography (Android Lens Blur, iPhone Portrait Mode), panorama construction (Google Photo Spheres), face detection, expression detection (smile), Snapchat filters (face tracking), Google Lens, Night Sight (Pixel) Medical imaging: CAT/MRI reconstruction, assisted diagnosis, automatic Pathology, connectomics, Al-guided surgery “Media: Visual effects for film, TV (reconstruction), virtual s Ports replay (reconstruction), semantics-based auto edits (reconstruction, recognition) ‘surance: Claims automation, Damage analysis, Property inspection One must Point out that even all these advancements in CV, Aland the area of computer vision still Need to tackle the basic problems associated wii unawareness and lack of explainability. ith it such as ethics, bias, riskTeachable Snake is an interactive web game powered by au beta version of Teachable Machine 2 and React.js, inspired by Webcam Pacman project ‘The idea is that instead of using physical buttons to contro! the game, everyone can draw a black arrow on a piece of white paper as controller, and move the snake by turning the paper in different directions in front of the webcam. To recognize the paper controller pointing at different re Stoved ate retain Wi 500 photos in each arrow direction (up, down, right and left) an inane With the be, version of Teachable Machine 2. With the API endpoint published, lel can be lig in the React.js app a Scan QR code or visit: htips://experiments.withgoogle.com/teachable-snake We all know that YouTube is essentially the second largest search engine with hundreds of hours of videos uploaded every minute and billions of videos being watched daily, The internet consists of texts, images and videos. We have been able to index and search texts since long and algorithmically as well, it is fairly straightforward. For images Particularly, we need to analyze what the image contains and for that we need Computer Vision. Over the years, meta tags and descriptions have been the best bet to describe images but with CY, it is possible to literally “see” and “interpret” the images and know what they contain, ee | a a) vate (yo ; Saati i 48 FallCourse Laieo o> il 6 Sascenens Be, Feecony Me Sey Pe SRE Serrercrnmay tone To rf. = OQ ere pene inn ty tts Brn ee San conch yao *2se00008 ; Iys see the process of how we typically visualize and interpret vision: can dese the contents of an image we see once : summarize an image or video that we have only seen once without any pro ser Ps asa wean recognize 2 face that we have seen before This is the bare minimum expectation from Computer \ difficult, right? n as well. Doesn’t sound very At an abstract level, the goal of computer vision problems is to use the observed e data to infer something about the world. —Computer Vision: Models, Learning, and Inference, 2012 mn is sometimes confused and used interchangeably with Image Processing. is defined as the process of creating a new image from an already z ge. ‘either simplifying or enhancing the content in some way. It is basically a Signal processing and is not concerned with interpreting the content of vision is to extract useful information from images. This has challenging task; it has occupied thousands of intelligent er the last four decades, and despite this we are still far d a general-purpose “seeing machine”. —Computer Vision: Models, Learning, and Inference, 2012vou can take # took et Google class Neve * © san OR code or visit Xe , a oh See the magic of Hololens Mixed Reality: ee Scan OR code or vis ener ‘See how VR can save lives by doing safety trainings: Scan OR code or visit NeuralCandy is a machine that dispenses candy when the user di splays the Fequested in front of the trained image model NeuralCandy combines image classifier and sugar highs in one delicious Android Things § The application asks for a random image to be placed in front of the camera module matches the request, then the motor of the candy dispenser is activated to release the "uses the TensorFlow Lite inference library for Android to locally classi against the pre-trained ImageNet model. This model is good at recogi it was trained with. You can use a smartphone to search on Google f9 image and put it in front of the Pi camera. The Raspberry Pi 3 model ge processing and the motor for the candy release. AptiFiciAL IM> a a ee al Language Processing (NLP) ature : fl wi ve seen earlier, unstructured data such as random comments, images, video dio contain a lot of information. Earlier, it was complicated and expensive to i we of it but with the advent of CY, it is easy to take care of images and videos. Ss! : ° oe t about text and audio? Here’s when Natural Language Processing (NLP) comes 5 we ha} put wha’ into play! - nip is all about the algorithms to understand and interpret speech and text which is atural language data. It could be any language, can be a sentence or a paragraph, NLP an scrape through it and make sense. traditionally, humans interact with each other in natural language but when it comes to interacting with machines, we use programming languages and software codes. What fun would it be if we could talk to machines in the same way as we talk to other people? NLP makes it possible. a8 The idea of a talking machine goes back nearly 200 years when Christian Kratzenstein built the first speaking machine in 1779 that could produce 5 long vowel sounds. In 1954, a machine was used to translate 60 Russian sentences into English using an IBM 701 mainframe computer. This success is largely seen as the beginning of modern Natural Language Processing (NLP). NLP is the heart of the modern Al input/output systems. The ability to understand natural language provides two important capabilities. The first is the ability to comprehend unstructured data, such as the 2.5 million peer reviewed papers published every year. The second is to mimic human conversation. Some of the well-known applications of NLP we use in our everyday life are: * Translation tools such as Google Translate, Microsoft Translator * Document Processors such as Microsoft Word and Grammarly that employ NLP to check grammatical, semantic and plagiarism to check accuracy of texts. Standard Interactive Voice Response (IVR) applications used in call centres to handle Support queries Personal assistant applications such as Google Assistant, Siri and Alexa.oan (NLU) systems is to figure out the stories and so on. Systems aima, mn f information: The goal of Natural Language Understanding of language inputs: the words, sere ies mic Problem use a combination of three differe' iptions. r 5 or descrip’ Pragmatics: Contextual information such as list ings can be combined, Semantics: The meaning of words and how those rene it types of language Syntax: The structural relationships among aie : “lem ents such as nouns, verbs, adjectives, prepositions and p! verts a full. T alled inshorts con) length A new article summarization app ¢ line and a ca news article to a 60-word news brief along with a headli 0 image (Source: £7) There is an a merging of CV and NLP in an application called WordLens. Have a look at it: Scan QR code or visit: Baie You can use the Apache Open source framework Apache OpenNLP for your ‘Symtanthae nn Pes *0 read and understand human language in is SymtaxNet, an open-source neural network framework for TensorFlow, mguage Understanding (NLU) systems, m/2017/030n-upg ESsenmiacs oF Aarieicia:y Se el ee | sovening 218012 © 319M 7478 views Forbes Earnings Preview: Chesapeake Energy By Narrative Science Conmaron ‘Analysts axpect decreased profit for Chesapeake Eneray (CHK) when the ‘Company epors ts fourth quarter results on Tuesday, February 21,2012 ‘Athough Chesapeake Energy reported profit of 70 cents» year ago, the ‘consensus estimate cals fr earings per share of $8 cents. What to Expeet: ‘verte past three months, the consensus estimate has fallen from 62 cents For the fiscal year, analysts ae expecting earnings of $280 per share. Revenue is expected to be $2.04 bilion fr the quarter, 53% high vyeor-earlir total of $1.98 bilion. For the year, revenue is proje at $11.87 bilion. | Trends to Watch (Source: Forbes.com) Fig. 1.14: An article written entirely by Quill, an advanced natural language generation platform. Scan QR code or visit: https://narrativescience. com/quill/thm openAl’s GPT3 language nths of training and tons of data on ay. i en could be readin; iy computers but once that’s done it’s easy tO use. 59 i it, you can fix : oes " 7 but not be sure who has written it. with a human writing " SOuntapi, : ble? on that person but with a machine, who is really account Do you know that a machine Jearning algori trained to predict text. It takes mo i car is fed with Consider the case of self driving cars. Imagine 2° auto car ine aM algort, to save the passengers sitting in the cer over people outside 4 7 Of acciden, otherwise no one will buy the car. During collision, to Save OMe PASSENEET sitting ing, . “di s the car, the car runs over a grouP of six people. This makes self-driving car Social, unacceptable, especially to those who cannot buy it- uture, how to make their decision-mak ble and certain under all circumstan: t machines and soft chines are to take over in the fi predicta If-improving independen So, if Al and mai reliable? How to make their decisions What will make these autonomous, se trustworthy? are This is where ethics come into picture. Ethics are loosely defined as a set of mora uiding actions of an individual or a group, helping to determine what is gy itself does not possess moral or ethical qualities, it needs to be When designed and tested well, it arrives at predictable outputs f rules or decision paths principles, g! or right. Since technolo; fed with human ethics. for predictable inputs via such a set 0 allenges here. First, how does the team of developers determin or But there are two chi what is a good or right outcome, and for whom? Is this outcome universally it good only for some? Is this outcome good under certain contexts or situ not under other conditions? Is it good against certain standards but not good ag others? These discussions, the questions and the answers ‘chosen’ by t eam are critical. The second challenge is that Al is an autonomous, self-learning and ving is means it does most of its decision-making itself based on its own analysis technology. Thi of data.ry of what All is Covered under Ethics yom) ics? of Al Eth! prejudice and Fairness 2 sparency Interpretability and Explainability a, Teh } actions OF Ab ‘ safety ; Human-Al inte : cybersecurity . vacy and Control raction and Wrong intentional Use 4 pata Pri Al Impact: 4, Job losses and Unemployment 2. Civil rights—Robot rights 3, Human-Human interaction Change 4. Economy that and if we don’t take specific actions to address it then we're just going to continue to perpetuate them or even make them worse,” OCR ORE a CRE aOR CELUI P= There is a fundamental debate now that Al will change the way our society works and it is very important to plan for such a society in advance before Al becomes too involved in our lives. Moral Issues: Self-Driving Cars Scenario 1: Let us imagine that we are in the year 2030. Self-Driving cars which are just a concept in today’s time are now on roads. People like us are buying them for ease and using them even for their daily commute. Of course, this car is expensive because of all the features that it has. Now, let us assume one day your father is going to office in his self-driving car. He is sitting in the back seat as the car is driving itself. Suddenly, a small boy comes in front of this car. This isso sudden that the car is only able to make either of the two choices: 1. Go straight and hit the boy who has come in front of the car and injure him severely, or a sharp right turn to save the boy and smash the car into a metal pole thus ing the car as well as injuring the person sitting inside.t the developer of th, stand tha e car With the help of this scenario, we need t© ine ~ar's algorithm. Thus, here the m2 through all such dilemmas while developing Me 55 what according to him/her rate of the developer gets transferred © the machin the selection made by the a "Rhy would have a higher priority and hence wou’ it tnere Was 7 other alternative ng if you were in the place of this developer aoe, Othe a cist why? situation, which one would you prioritize an sume that the cat should be . hit the boy who came in from. Id responsible forit? Why? Scenario 2: Let us now as: Considering this as an accident who 1 2. The manufacturing company; 3. The developer who developed the car’s algorithm; or 4. The boy who came in front of the car and got severely injured? erson to person and one must understand «3 jective and hence they The person who bought this car; Here, the choices might differ from p nobody is wrong in this case. Every person has a different persp' take decisions according to their moralities. Data Privacy s around data. Every company, whether big % of the datd The world of Artificial Intelligence revolve or small, is mining data from as many sources as possible. More than 70% hree years which shows how importa collected till now has been gathered in the last ¢ become in recent times. It is not wrongly said that Data is the new gold. Thi data has makes us think: Where do we collect data from? We need to understand that data, which is collected by various applications, is as smartphone users agree to it (by clicking on ‘allow’ when asked for permision by agreeing to the terms and conditions). But at the same time if one does not "@" to share their data with anyone, they can opt for alternative applications which have aeWhatsApp |s ge and keep your data private. For example, an alternative to a a a app which does not collect any data from us. But since WhatsApp Is more eee 4 used by most, people go for it without thinking twice. popula an' qe Key eans reattime ana expand the V6 ivacy int ponentially. components of Big Data are the three “V's”: Volume, Variety and Velocity. Volume ore data; variety facilitates more powerful and detailed analysis; velocity enables lysis and sharing. Streams of data from mobile phones and other devices olume, variety and velocity of information about our personal lives and bring Pr he spotlight. The ‘volume’ refers to the quantities of big data, which is increasing & we need to protect individuals against any adverse effects from the use of personal information in Al. We need to tackle the issues of predictive policing that could affect minorities OF 3 hiring algorithm that can impact a certain gender in the workforce. addressing algorithmic discrimination is the key point in handling data privacy. some of the key considerations while handing data privacy issues are: + Data stewardship requirements such as duties of fairness « Data transparency and rights of individuals to access information + Data governance rules through “privacy by design” + Rules on data collection and sharing “OSD i goon KS #1: #2: #3: #4: Embedding Proliferating Increasing Access control data privacy devices maintenance costs is difficult in many industries v | | #8: 47: #6: #5: | Along list of The ever-increasing Abad Getting visibility | regulations and scale of data data culture into all your data documentation | tofollow Fig. 1.15: Data Privacy