Training the model

In this lesson we will explore the concept of training the model through the data.

Let’s start with an example.

You have a coffee machine that doesn’t know how to make coffee in a known machine learning setting.

We would need to incorporate the instructions in the machine’s electronics or at least that’s how all

coffee machines I have seen work.

For instance the engineers may set the machines electronics to first grade 20 grams of coffee beans

second heat the water to 200 degrees Fahrenheit.

Third poor 100 milliliters through the ground coffee into the cup.

Now in a machine learning setting we won’t explicitly provide instructions to the machine.

Instead we just need to state our goals in the coffee machine case.

That would be produce a cup of coffee then we would let the machine work out the problem on its own.

The machine learning process is a kind of trial and error training.

The machine would try various combinations of grinding heating and pouring.

Most would not make sense.

The machine would try heating the water and pouring it before grinding the coffee resulting in a cup

of hot water or it may grind the coffee and pour the water without heating the water.

Anyhow after thousands of trials and errors the algorithm would train itself to reach the set goal every

time.

It could make a cup of coffee.

It is possible that it will learn to make the best coffee you have ever tried.

Much better than the one obtained by following a set of instructions.

That’s because it would have gone through so many more recipes than a human would ever be able to a

reasonable optimization algorithm would not try all combinations as there are usually inexhaustibly

many other options in the Kovi example.

If the coffee machine learns that grating the coffee has to go before pouring the water it would not

waste time attempting it in the wrong order.

We’ll talk more about optimization algorithm soon.

Don’t worry this example shows us why machine learning is so powerful it allows system to learn on their

own situations where humans cannot define a rigid set of rules for the computer to follow.

Even if we can define a set of rules an algorithm can probably provide a better one before wrapping

up.

Let’s explore another interesting instance self-driving cars.

Contrary to what many people think self-driving cars don’t follow rules such as avoid curbs.

Essentially they train on thousands of hours of footage of real people driving and learn how to mimic

them efficiently.

It is not the set of rules they know but the final goal the final goal fundamentally is to drive safe

and efficiently avoid curbs.

Don’t bump into other cars don’t go over the speed limit stop at red lights and so on.

OK now that we know training data is at the core of the process we can differentiate between different

types of machine learning.

Stay tuned and thanks for watching.