Original subtitles

Instructor: Hey, this is the last theoretical section

of the course.

It is about the first activity you want to do

when you start creating a machine learning algorithm.

Preprocessing.

Preprocessing refers to any manipulation

we apply to the data set

before running it through the model.

Everything we saw so far was conditioned

on the fact that we had already pre-processed our data

in a way suitable for training.

You've already seen some preprocessing.

In the TensorFlow intro, we created an npz file.

All the training we did came from there,

so, if you must work with data in an xl file,

CSV, or whatever,

saving it into an npz file

would be a type of preprocessing.

In this section though,

we will mainly focus on data transformations

rather than reordering as before.

What is the motivation for preprocessing?

There are several important points.

The first one is about compatibility

with the libraries we use.

As we saw earlier,

TensorFlow works with tent source

and not Excel spreadsheets.

In data science,

you will often be given data in whatever format

and you must make it compatible with the tools you use.

Second, we may need to adjust inputs of different magnitude.

Let's say we are 4x traders.

If one input we are working with

is the end of the day Euro/Dollar exchange rate,

it would be a value around 1.

However, if another input is the daily trading volume,

we would have values like 100,000 and higher.

Obviously, the orders of magnitude are quite different.

A linear combination of numbers

based on such different skills as problematic.

In purely mathematical terms,

a value of 1 is negligible regarding a value of 100,000.

As all the inputs are on an equal footing in a vector

or a matrix,

the algorithm is likely to ignore all values around 1.

These values essentially represent

the Euro/Dollar exchange rate itself,

so they are often more important than the volume of trading.

Obviously, something needs to be done to solve this issue.

A third reason is generalization.

Problems that seem different

can often be solved by similar models.

Standardizing inputs of different problems

allows us to reuse the exact same models.

Sometimes there are cases

when we can even reuse already trained networks.

Imagine that!

You have trained a model previously,

you face a new problem, you test your model,

and it works like a charm.

That's not unusual in machine learning.

In the next few lessons,

we will focus on these concepts

and introduce several pre-processing techniques.

Thanks for watching.