AI, ML, DL and Everything in Between

Introduction

We live in the information age. Today, information spreads like a swarm of bees, who’s hive has just been hit by a rock, creating buzz… everywhere.

Let not the buzz, blur the line between Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL) and Data Science (DS).

1. Artificial Intelligence: AI is the superset which encapsulates Machine Learning and Deep Learning. In generalized terms, any machine which mimics human behaviour can be categorized as an Artificial Intelligence machine.
2. Machine Learning: A subset of AI and a superset of DL. The objective of an ML model is to learn from the patterns emerging in the past experiences or data and to map future decision.
3. Deep Learning: DL, mimics the functionality of human brain using neural network architectures, which are anatomically similar to the structure of the human brain
4. Data Science: Applied Data Science could be best defined as a combination of all above technologies with business knowledge, to drive business objectives further.

Why Machine Learning?

Suppose you have a vast collection of pictures of different kinds of animals. Now you want to make a collage of only those pictures which are of Dogs. Then to classify each image as a picture of a Dog or of some other animal, you could take two approaches.

1. Heuristics Based Approach: You can select few Dog pictures, and then you can manually come up with a set of features which the Dogs in the pictures generally seem to possess.
2. ML Based Approach: You can feed the data and output labels to the machine for a plenty of pictures, so that the machine can analyse and learn all the possible features that should be used to classify a sample as either a Dog picture or as a picture of some other animal.

The problem with a heuristics based approach is, there are thousands of Dog breeds. Some are large, some are small, some have blue eye and others are just too similar to Foxes and thus, too difficult to manually distinguish. It is impossible to map their features manually. Thus, the set of features that you can manually identify would always be non-exhaustive. There will always be Dogs, which do not fit the description that you have manually prepared.

A machine, on the other hand, can process a large amount of data within seconds and, it can also distinguish the minutest of the differences. Thus, an ML based approach explores all possible pattern which can be used to map the input data to the target labels. An ML based approach, produces an exhaustive set of feature, and thus it is more complete.

Tasks of Machine Learning

Source: AlmaBetter

Supervised Learning

Whenever the main problem at hand is to map the given data to an outcome or a label which could also be called a dependent variable, then it is a Supervised Learning Problem.

1. If the dependent variable at hand is of a continuous type, then it is a Regression Problem. Forecasting the price of a Car based upon a list of its given features could be an example of a Regression Problem.
2. If the dependent variable is of categorical type, then the problem at hand is a Classification Problem. Classifying animal pictures as ‘Dog picture’ or ‘not a dog picture’ could be an example of a Classification Problem

Unsupervised Learning

When, a target label/dependent variable is absent from the problem, then it is an unsupervised Learning problem. Here, we are not required to determine labels for a given data, but we are required to find out clusters in our data. A cluster is, nothing but, a compartment/group or a part of our data which shows similar characteristics. Based on what action people take when a particular recommendation is shown to them on Netflix, they can be clustered into different types of users. Going forward, Netflix could create tailor made recommendations for each user type. This is how powerful Clustering Algorithms are when it comes to understanding customer user behaviour.

Reinforcement Learning

Source: AlmaBetter

A reinforcement learning model, has an agent which takes a certain action in an environment with a certain initial state, S0. For the action taken by the agent, the environment provides it feedback or a Reward. After receiving the feedback, the Agent changes the state to, say S1. With an optimized state, the agent continues to produce action till the best possible optimization is not achieved for the state of the algorithm.