Prior to general machine learning, if we wanted to, for example, build a spam filter, we could start by compiling a list of words that commonly appear in spam. The spam detector then scans each email and when the number of blacklisted words reaches a threshold, the email would be classified as spam. This is called a rules-based approach, and is illustrated in the following diagram:

The problem with this approach is that once the writers of spam know the rules, they are able to craft emails that avoid this filter. The people with the unenviable task of maintaining this spam filter would have to continually update the list of rules. With machine learning, we can effectively automate this rule-updating process. Instead of writing a list of rules, we build and train a model. As a spam detector, it will be more accurate since it can analyze large volumes...

There are several distinct types of learning tasks that are partially defined by the type of data that they work on. Based on this, we can divide learning tasks into two broad categories:

• Unsupervised learning: Data is unlabeled so the algorithm must infer a relationship between variables or by finding clusters of similar variables
• Supervised learning: Uses a labeled dataset to build an inferred function that can be used to predict the label of an unlabeled sample

Whether the data is labeled or not has a predetermining effect on the way a learning algorithm is built.

One of the main drawbacks to supervised learning is that it requires data that is accurately labeled. Most real-world...

It is important to remember that an image detection model does not see an image but a set of pixel color values, or, in the case of a spam filter, a collection of characters in an email. These are raw features of the model. An important part of machine learning is feature transformation. A feature transformation we have already discussed is dimensionality reduction in regard to principle component analysis. The following is a list common feature transformations:

• Dimensionality reduction to reduce the number of features using techniques such as PCA
• Scaling or normalizing features to be within a particular numerical range
• Transforming the feature data type (for example, assigning categories to numbers)
• Adding random or generated data to augment features

Each feature is encoded on to a dimension of our input tensor, X, so in order to make a learning model as efficient...

Choosing a model representation is an important task in machine learning. So far, we have been referring to models as black boxes. Some data is put in, and, based on training, the model makes a prediction. Before we look inside this black box, let's review some of the linear algebra that we will need to understand deep learning models.

Linear algebra is concerned with the representation of linear equations through the use of matrices. In the algebra taught in high school, we were concerned with scalar, that is, single number, values. We have equations, and rules for manipulating these equations, so that they can be evaluated. The same is true when, instead of scalar values, we use matrices...

As the name suggests, ANNs are inspired by their biological counterpart, although the reason is, perhaps, misunderstood. An artificial neuron, or what we will call a unit, is grossly simplified compared to a biological neuron, both in terms of functionality and structure. The biological inspiration comes more from the insight that each neuron in a brain performs an identical function regardless of whether it is processing sound, vision, or pondering complex mathematics problems. This single algorithm approach is, fundamentally, the inspiration for ANNs.

An artificial neuron, a unit, performs a single simple function. It adds up its inputs and, dependent on an activation function, gives an output. One of the major benefits of ANNs is that they are highly scalable. Since they are composed of fundamental units, simply adding more units in the right configuration...

We have covered a lot of material in this chapter. Don't worry if you do not understand some of the mathematics presented here. The aim is to give you some intuition into how some common machine learning algorithms work, not to have a complete understanding of the theory behind these algorithms. After reading this chapter, you should have some understanding of the following:

• General approaches to machine learning, including knowing the difference between supervised and unsupervised methods, online and batch learning, and rule-based, as opposed to model-based, learning
• Some unsupervised methods and their applications, such as clustering and principle component analysis
• Types of classification problems, such as binary, multi-class, and multi-out classification
• Features and feature transformations
• The mechanics of linear regression and gradient descent
• An overview of...