Hands-On Differential Privacy

Appendix F. Machine Learning

This appendix highlights the basic terminology of machine learning, outlines several varieties of machine learning, and summarizes the use of gradient descent to learn parameters.

The process of fitting a machine learning model is called training. Given a function with some unknown parameters, the training process tries different parameter values until it finds the values that allow it to most closely match the data. To understand how accurate the model is, you perform testing; this is the part of the process where you use your parameters to make predictions based on the input data and compare the predictions to the true values. If the tests show that your model is accurate, you’ve trained the computer to learn a model about the data. The trained model can be used to make decisions or predictions.

Supervised Versus Unsupervised Learning

Machine learning is often broken into two broad categories: supervised and unsupervised. Supervised learning describes a scenario where the data comes with labels to learn from. These labels aid the model in learning patterns between the different categories. For example, Table F-1 shows a data set of fruit measurements.

Table F-1. Fruits in a grocery store

Height (cm)	Diameter (cm)	Type
6.2	6.7	Apple
7.0	6.9	Orange
9.0	6.2	Pear
6.3	6.5	Apple
7.1	6.8	Orange
6.4	6.8	Apple
9.1	6.1	Pear
9.2	6.2	Pear
7.2	7.4	Orange

Now imagine you are given a task—you know the height and circumference of a fruit ...