machine learning decoded greatlearning
machine learning decoded greatlearning

If they be two, they are two so
As stiff twin compasses are two;
Thy soul, the fixed foot, makes no show
To move, but doth, if the other do.
 And though it in the center sit,
Yet when the other far doth roam,
It leans and hearkens after it,
And grows erect, as that comes home.
Such wilt thou be to me, who must,
Like th’ other foot, obliquely run;
Thy firmness makes my circle just,
And makes me end where I begun.
I love John Donne’s poetry. I always wondered if I would ever be able to write like that or come up with an analogy as brilliant as the prongs of a compass compared to the two souls in love. While I stroll around doing my usual business, I trust that there will come a day when some brilliant data scientist, hopelessly in love, and inclined to poetry will come up with an algorithm to find another analogy that will do John Donne proud and give me something to rave about. Until then, the good Samaritan that I am will help someone become a data scientist to inch closer to making it a reality.

So, What Exactly is Machine Learning?

Machine Learning is a subset of Artificial Intelligence such that a computer is able to learn on its own without explicit programming code. It uses statistical techniques to enable the system to ‘learn’ better and make future decisions and predictions based on the million data entries it has already evaluated. It grows in sophistication and accuracy and improves with every new experience or exposure (to the dataset.) A machine learning algorithm gains more and more ‘experience’ every time it processes a new data set.
The advanced applications of machine learning are known to mimic the human brain. As a baby, you start to learn different things. Your mother tells you that “this is a ball” hundreds of times before you can register it implicitly. Then comes an orange and as a child you think it is a ball (it is small, green, and round) until you grow up to add other variables like edible/non-edible, living/thing, squishy / firm etc. to decide if it is a ball or an orange. Machine learning as a branch of computer science enables a system to learn the exact same way.
Apple’s Siri, Google’s Personal Assistant, and Amazon’s Alexa are all examples of Machine Learning. Driverless cars, sentiment analysis, robotics are more advanced applications where machine learning is the prime catalyst. Corporate giants like Uber and Amazon are using machine learning at an unprecedented level of sophistication. While Uber schedules its rides, develops effective geo-mapping techniques, identifies hotspots for demand, and determines surge pricing; Amazon uses machine learning in several of its products like Alexa, Polly, Prime Air, recommendation engines on its e-commerce site, etc.

Types of Machine Learning

Machine learning is of various types and anyone interested in pursuing an artificial intelligence course or a machine learning certification must know what they are dealing with. A career in artificial intelligence and machine learning will always be super exciting but seldom a walk in the park. Let’s get started:

Supervised Learning

Supervised Learning infers a function based on the labelled training data fed to the system which is then used for mapping newer data. Supervised learning generalizes the learning algorithm from training data to unseen data to draw conclusions. The input data is X, and the output is Y, so Y = F(X). This exact rule is then applied to unseen data. There are several techniques of supervised machine learning algorithms such as linear regression, logistic regression, decision trees, and multi-class classifications, and support vector machines. All of this may sound like Greek in the beginning, but here are a couple of examples explaining supervised learning:

  • – One of the classic examples of supervised learning is the spam folder created with your Gmail or icloud account. A machine learning algorithm classifies whether an email is spam or not given a particular set of attributes. These attributes/variables could be whether you have moved similar emails (emails from the same user, same subject, etc.) into the spam folder or Google’s own algorithm of finding cues in the subject and mailer content that classifies as spam. This is an example of classification.
  • – Another one is as simple as choosing your office wardrobe. Supervised learning can classify if a dress or outfit falls into the “office-wear” category or not by determining the color, type, fabric, hemline, etc. Isn’t this how we buy clothes? Most of our choices and ‘fashion-sense’ can be defined by a set of variables and rarely do we ever experiment outside the box. You will be more surprised at how random it may seem, and yet how patterns are easy to identify if we get down to it.
  • – If these are too generic for you, let us consider a real business problem of credit score. A common scenario is a bank wanting to know if a particular account holder is a good case for a loan or a credit card. They evaluate so by giving a credit score to the customer. Millions of data records from past transactions and customer history are fed to the system for it to calculate if a new candidate is eligible for a loan.

Unsupervised Learning

In the absence of a training data set and direct input (X), no corresponding output is known beforehand. This is unsupervised learning. It is based on finding hidden trends and patterns that may not have a set output. Think of any classic business problem. Not all cases will work on supervised learning. The need for unsupervised learning is more a consequence of symptomatic data that needs to be evaluated in order to figure out a potent solution. Several techniques are used for unsupervised learning such as k-means clustering, association rules, principal and independent component analysis, and apriori algorithms. Unsupervised Learning has an alchemical feel to it!
Let’s look at a couple of examples:

  • – Suppose that it is your first day at a new office and you know no one. Since you have no prior information about any of them, you would start by classifying them based on gender, age, experience, demeanor, appearance, gait, behavior, etc. This is a classic example of unsupervised learning where you didn’t know the people or the attributes that define them but at the end of the day, you have some idea about who you are likely going to be friends with, or work well together. You would be able to easily spot clusters of people who have lunch together, who hang out together, take breaks together, etc. If you need more information about a person (Person A), you will identify with ease the person who would most likely know Person A the best.
  • – Another example is sentiment analysis of informal textual or non-textual communications on the internet. Communication can be textual like a tweet or an online forum and it can be non-textual such as video, audio, and images. Unsupervised learning makes it possible to understand the sentiment behind an image or a tweet by finding structures or relationships between the different inputs here. Various studies have found the patterns drawn from unsupervised learning can be more accurate than other machine learning methods outperforming the myths around sentiment analysis and providing reliable solutions.

Reinforcement Learning

Reinforcement learning has an input, a hypothesis of an outcome, and the grade for output. Inspired by behavioral psychology, it uses observations from the environment to take actions maximizing rewards and minimizing risk. A reinforcement learning algorithm (agent) learns from the environment iteratively until it has exhausted all possible states. ‘Ideal behavior’ is determined in a specific context by measuring performance based on reward feedback. The end goal is to maximize the reward feedback in each case. Q-Learning, Temporal Difference (TD), Deep Adversarial Networks are some of the common reinforcement learning algorithms. Self-driving cars, robotics, and games such as chess, etc. are some of its popular applications.

  • – It mostly applies to gaming where it rates the moves made by a player and learns accordingly. Imagine yourself playing chess online. The computer doesn’t know your next move. But once you move a piece, reinforcement algorithm will help it react with the environment to find its most rewarding move as a response to yours. The system will make a move according to the initial state and the action performed to change it and determine positive or negative reward to decide.
  • – Self-driving cars is work in progress by companies like Uber, GM, Ford, etc. “By 2021, Ford hopes to have a self-driving vehicle with ‘no gas pedal’ and ‘no steering wheel,’ with no need for the passenger to take control “in a predefined area,” Ford Motor CEO, Mark Fields told CNBC at the Detroit auto show. He further added, “In our industry, the word autonomous is being used very, very liberally. There’s different levels of autonomy. The question that should be asked when a company says they’re going to have an autonomous vehicle … is at what level.” (SAE International determines the capability of a self-driving vehicle on automation from zero to five – level five meaning complete automation).

Applications of Machine Learning

Some of the most common applications of Machine Learning are:

  1. Natural Language Processing – NLP is critical for bridging the gap between machines and humans. Its complexity arises from the fact that it takes years for a human being to build command over a language and yet it is so difficult for them to stick to rules and semantics. Human Language is ambiguous wrapped in layers of context, tone, and subtleties making it excruciatingly painful to define with an algorithm. But as with all matters, homo sapiens are nowhere near giving up. Constant work on NLP is in progress where scientists are trying to capture the true meaning, context, and emotion behind a sentence by collecting metadata.
  2. Medical Outcomes Analysis – Medical practitioners will be able to predict the lifespan of those suffering from diseases with more and more accuracy. Whether it is drug discovery or management, medication or best course of treatment, disease prevention, or reducing fraud and abuse in the healthcare sector, machine learning is already making waves. McKinsey Global Institute estimates that applying machine learning techniques to better inform decision making could generate up to $100 billion in value based on optimized innovation, enhanced efficiency of clinical trials and the creation of various novel tools for physicians, insurers, and consumers.
  3. Banking and Fraud Detection – One of the key areas where the finance sector is benefitting from machine learning happens to be fraud detection. Machine learning is enabling corporates to identify fraudulent transactions thanks to the cashless transactions and trail data available for every account holder. It won’t be long before robots also start recommending investment schemes based on an account holder’s financial history. Citibank has collaborated with Portugal based fraud detection company Feedzai that works in real-time to identify and eliminate fraud in online and in-person banking by alerting the customer.
  4. Marketing – According to Forbes, “84% of marketing organizations are implementing or expanding AI and machine learning in 2018 and 75% of enterprises using AI and machine learning enhance customer satisfaction by more than 10%.” Sales and Marketing are experiencing the power of machine learning firsthand by accurately measuring the impact of their campaigns in real time, engaging the right audience, bringing in quality leads, and overall reducing costs to maximize profits.
  5. Robotics – More as robot learning, machine learning is enabling robots to increasingly behave like humans but process information faster like supercomputers. From drones to self-driving cars, performing surgeries to manufacturing products at lightning fast speed, machine learning is being used extensively to process physical data in dynamic environments, learn by imitation, and take data-driven decisions accurately. According to Harvard Business Review, “The Age of Smart, Safe, Cheap Robots Is Already Here: As technology has advanced and robot production has scaled up, costs have fallen by about 50% since 1990 — while U.S. labor costs have risen 80%. In China, manufacturing wages have risen five-fold just since 2008 as employers have chased workers eager to switch jobs for better pay.”

Hope it helped thou understand machine learning, its applications, and its types. Its applications may seem humdrum but take a closer look and all the stuff from sci-fi movies will come to life. As for me, I am off to reading more metaphysical poetry now!



Please enter your comment!
Please enter your name here

seventeen − ten =