Three Pillars Of Machine Learning- Supervised, Unsupervised and Reinforcement Learning.

What Is Machine Learning ?

Machine Learning is a branch of Artificial Intelligence (AI) and as the name suggests it is an artificial human brain which tries to imitate how the human responds to a particular situation irrespective of weather he is accustomed to that situation in the past or not. No system is perfect and so due to the alias 'Artificial', Machine Learning comes handy with the errors in decision Making. So, there is always a scope of improvement in the field of Machine Learning.

Three Pillars Of Machine Learning- Supervised, Unsupervised and Reinforcement Learning.

Before Analysing any further, consider the following Example-

Suppose, you have become the Parent for the first time and you have decided to order a cradle from Amazon. Obviously, the first thing you'd do is to assemble the cradle. For this, you will search for the user manual right? You will go step by step and will try to follow the instructions. Now, suppose if you have yet not assembled the cradle and your maid has swept off the instruction manual, then you are left with no choice but to do the guess work, basically do the hit and trial to assemble the piece. You will start by first grouping the similar pieces at a place an then probably start of the assembling process.

The above scenario is similar to how machine Learning works. With the data and problem statement in mind, a programmer choose how he wants to train the algorithm using a particular learning. The different types of learning are supervised, unsupervised and reinforcement learning. Lets talk about each of them in detail.

Supervised Learning

Consider that you are learning to drive a car. You are taking daily instructions from your driving teacher who is supervising if you are driving correctly or not. This is a similar situation of supervised learning. You have a supervision/guidance on you. Pertaining to the particular features, you are labeled whether you belong to A category, B category and so on (Can also be extended to continuous data). Thus, input data is labelled to an output and a model can learn from these instances. 

Now that you have a decent idea of what a Supervised Learning is, let us dig deep into the types of Supervised Learning.

The two types are :-

1) Classification

2) Regression

Classification problem trains the algorithm in such a way that it classifies the test data into one of the  class or group. Suppose, if we have trained our model which predicts weather the temperature is hot or cold tomorrow based on the past patterns or learnings, it comes under Classification Supervised Learning.

Regression Problems are for continuous data example- house price, age, weight etc. Take the case to predict the temperature. Here, despite weather tomorrow will be hot or cold, our classifier will try to predict the numerical temperature of tomorrow based on the past learnings.

 

Unsupervised Learning

Now that we have learnt that supervised learning has the labeled data which is fed to train our classifier. The major difference between between supervised and unsupervised learning is that there is no cleaned labeled data in unsupervised learning.

No Cleaned labelled data? Confused?

Let me articulate the above saying. Unsupervised Learning is a self learning algorithm that aims to find the patterns or useful information from the unlabeled data.

Still Confused ?

In Unsupervised learning, a model receives the data without guidance. Boom! Did you relate with the cradle example in the starting where the user manual was swept off by the maid and the person had no choice but to set the cradle on the basis of some patterns? Yes, I am sure you must have. Similarly, in Unsupervised learning, a model gets the dataset without labels. Also, you don't have any clue that what do you expect from the data yet. You might be thinking that there might be a relationship between the data , but it is impossible to manually find the relationship between the data especially when the data is big data. What actually happens then? Well in that case grouping is done on the basis of patterns and the model makes the comparisons to predict the output.

Now let us try to understand unsupervised learning with a real life example.

Consider that we are given unlabeled dataset which has different types of animals. How will my model predicts weather my animal is cat, bird or a fish? If my model gets the information that an animal has beak, feathers, wings, etc it might be a bird. Similarly if a animal has gills, lives in water, tail, it might be a fish, or if an animal has fluffy fur, curly tail, floppy ear, it might be a cat.

Hence, based on this information, our model might be able to distinguish between different types of animals.

Reinforcement Learning

People generally gets confused in reinforcement learning. Lets me try to spot some light on this concept. To be specific, in reinforcement learning, algorithms learn to react to an environment on their own.

To be specific, there is a start and an end state for an agent (AI driven system). There are multiple ways to reach from start point to end point, like in a maze. By applying hit and trial, the algorithm learns itself to reach an end pint. Popular examples of reinforcement learnings are self driving cars, automatic vacuum cleaner., etc. Every time, an agent/machine takes a correct step, it is appreciated and otherwise penalised for the wrong step.

Lets go through a journey where a child is taking his first step. What are the observations that a child makes?

1) Observing others and copying them.

          2) Trying to stand with the help of a wall.

          3) Trying to balance and remain still.

          4) While Balancing, deciding which leg to put forward.

It seems a difficult task for a child, right? But for an adult it is a cakewalk. From the very first step till date, the adult has been awarded and penalised for the corresponding act and thus it has made him expert in walking.

Now putting it in a nutshell, here child is an agent who is trying to work through the environment (floor). Every time he takes a step, he is awarded but will not receive an appreciation if he is unable to walk. This is the overall gist of the reinforcement learning.

I think with the help of this blog I have made myself clear that -

1) Supervised learning is a learning with guidance ( labelled data).

2) Unsupervised learning is a learning without guidance.

3) Reinforcement learning is a learning where a machine or an agent interacts with an environment, and performs actions on the basis of hit and trial.