Thread by @ammaryh92

The term Machine Learning sounds mysterious and confusing to a lot of people especially beginners.
In this thread, I will try to explain how does a machine learn, and why do we even need machine learning?
🧵👇

In pre-machine learning era, we had what is called "rule-based systems".
This basically means that we provide a machine with a bunch of instructions on how to perform a certain task.
For example, if we need to write a function that returns the square of a number.

With rule-based system, this is very easy.
1. First we define a function called Square, for example.
2. Square function takes X as an input, where X can be any number.
3. Square function multiplies X by itself (X **2).
4. Square function returns the result to the user.

So in the previous example, we had to write the rules that map from the input X to the output Y, where Y = X**2.
We did not need machine learning for that because the mapping rules were relatively easy to understand and write.
Now consider the next problem.

We have a bunch of information about a particular real estate in New York regarding its location, number of rooms, population of the area, etc.
We need to write a function that takes these information as an input and can predict the price of the property.

As you can probably see, using rule-bases systems, it is almost impossible to come up with rules that can map the input features to the output target value. This is where machine learning shines.
The amazing thing about machine learning is that we do not have to write the rules.

All we have to do is to feed both the input information (X) along with the expected output values (Y) to the machine learning algorithm, and it will figure out the rules that maps X to Y itself. Most of the times, we do not even know what the rules are nor do we care.

In this example, we used what is called supervised machine learning, because the algorithm learns by taking both the input features X and the expected solution Y as inputs.

Because machine learning usually handles much more complex problems, we usually have thousands and sometimes millions of rules, called parameters, that maps the input to the output.
Using some algorithms, the machine can learn these rules itself, hence the name machine learning.

Machine learning is used for problems where the current solutions require a lot of tuning and maintenance.
Machine learning is also used for problems that are too complex to tackle using rule-bases systems.
However, machine learning is not the answer to everything.

The learning process is guided by different machine learning algorithms, chief among which is Gradient Descent.
The basic principle behind gradient descent was first published in 1986 in the amazing paper "LEARNING INTERNAL REPRESENTATIONS BY ERROR PROPAGATION".

a link to the paper
apps.dtic.mil/sti/pdfs/ADA16…
That is it for this thread.
In a future thread, I will discuss how exactly gradient descent works to learn the parameters that guide the mapping from the input X to the output Y.

If you find this thread to be useful, kindly consider retweeting it.
And for more content about machine learning, AI, and data science, please follow me @ammaryh92.