33. Standardizing Output

Numerical vs. Categorical

Let's learn a little more about labels. After visualizing the image data, you'll have seen that each image has an attached label: "day" or "night," and these are known as categorical values.

Categorical values are typically text values that represent various traits about an image. A couple examples are:

• An "animal" variable with the values: "cat," "tiger," "hippopotamus," and "dog."
• A "color" variable with the values: "red," "green," and "blue."

Each value represents a different category, and most collected data is labeled in this way!

These labels are descriptive for us, but may be inefficient for a classification task. Many machine learning algorithms do not use categorical data; they require that all output be numerical. Numbers are easily compared and stored in memory, and for this reason, we often have to convert categorical values into numerical labels. There are two main approaches that you'll come across:

1. Integer encoding
2. One hot-encoding

Integer Encoding

Integer encoding means to assign each category value an integer value. So, day = 1 and night = 0. This is a nice way to separate binary data, and it's what we'll do for our day and night images.

One-hot Encoding

One-hot encoding is often used when there are more than 2 values to separate. A one-hot label is a 1D list that's the length of the number of classes. Say we are looking at the animal variable with the values: "cat," "tiger," "hippopotamus," and "dog." There are 4 classes in this category and so our one-hot labels will be a list of length four. The list will be all 0's and one 1; the 1 indicates which class a certain image is.

For example, since we have four classes (cat, tiger, hippopotamus, and dog), we can make a list in that order: [cat value, tiger value, hippopotamus value, dog value]. In general, order does not matter.

If we have an image and it's one-hot label is [0, 1, 0, 0], what does that indicate?

In order of [cat value, tiger value, hippopotamus value, dog value], that label indicates that it's an image of a tiger! Let's do one more example, what about the label [0, 0, 0, 1]?