We have already seen what is meant by ML. In this section, we will focus on several terminologies such as supervised learning and unsupervised learning, and we will be taking a look at the steps involved in a standard ML workflow. But you may ask: why ML? We are supposed to learn about the applications of DL in this book. We just learned that DL is a type of ML only. Therefore, a quick overview of the basic ML-related concepts will certainly help. Let's start with several types of ML and how they differ from each other. 

ML encompasses a multitude of algorithms and topics. While every such algorithm that makes up an ML model is nothing but a mathematical computation on given data, the form of data that is provided and the manner of the task to be performed on it might hugely vary. Sometimes, you might want your ML model to predict future house prices based on the data of previous house prices with respect to details of the house such as the number of rooms and number of stories it has, and at other times, you might want your ML model to learn how to play computer games against you. You can easily expect the input data for the first task to be in tabular format, but for the second example, you might not be able to come up with the same. Hence, ML algorithms branch into three major categories and another form that derives from them, based on the input data they receive and the kind of output they are supposed to produce, namely, the following:

• Supervised learning
• Unsupervised learning
• Reinforcement learning
• Semi-supervised learning

The following diagram captures the three major types of ML, along with the hybrid form as a fourth type, and a very brief summary on each type:

You may have heard of the fourth form of ML—semi-supervised learning, which fuses both the worlds of supervised and unsupervised learning.

Let's now understand these types of ML in greater depth, according to how they function and the types of problems they can be used to solve. 

In this form of ML, the algorithm is presented with a huge number of training samples, which contain information about all of the parameters, or features, that would be used to determine an output feature. This output feature could be a continuous range of values or a discrete collection of labels. Based on this, supervised ML algorithms are divided into two parts:

• Classification: Algorithms that produce discrete labels in the output feature, such as normal and not normal or a set of news categories
• Regression: When the output feature has real values, for example, the number of votes a political party might receive in an election, or the temperature of a material at which it is predicted to reach its melting point

Most ML enthusiasts, when they begin their study of machine learning, tend to familiarize themselves with supervised learning first due to its intuitive simplicity. It has some of the simplest algorithms, which are easy to understand without a deep knowledge of mathematics and are even derived from what mathematics students learn in their final years at schools. Some of the most well known supervised learning algorithms are linear regression, logistic regression, support vector machines, and k-nearest neighbors.

Unsupervised learning presents itself in scenarios where the training samples do not carry with them output feature(s). You could wonder then, what are we supposed to learn or predict in such situations? The answer is similarity. In more elaborate terms, when we have a dataset for unsupervised learning, we're usually trying to learn the similarity between the training samples and then to assign classes or labels to them.

Consider a crowd of people standing in a large field. All of them have features such as age, gender, marital status, salary range, and education level. Now, we wish to group them based on their similarities. We decide to form three groups and see that they arrange themselves in a manner of gender—a group of females, a group of males, and a group of people who identify with other genders. We again ask them to form subgroups within those groups and see what people make groups based on their age ranges—children, teenagers, adults, and senior citizens. This gives us a total of 12 such subgroups. We could make further smaller subgroups based on the similarity any two individuals exhibit. Also, the manner of grouping discussed in the preceding example is just one among several manners of forming groups. Now, say we have 10 new members joining the crowd. Since we already have our groups defined, we can easily sort these new members into those groups. Hence, we can successfully apply group labels to them.

The preceding example demonstrates just one form of unsupervised learning, which can be divided into two types:

• Clustering: This is to form groups of training samples based on the similarity of their features.
• Association: This is to find abstract associations or rules exhibited between features or training samples. For example, on analyzing a shop's sales logs, it was found that customers buy beer mostly after 7 p.m.

K-means clustering, DBSCAN, and the Apriori algorithm are some of the best-known algorithms used for unsupervised learning.

Reinforcement learning (RL), is a form of ML wherein a virtual agent tries to learn how to interact with its surroundings in such a way that it can achieve the maximum reward from it for a certain set of actions. 

Let's try to understand this with a small example—say you build a robot that plays darts. Now, the robot will get a maximum reward only when it hits the center of the dartboard. It begins with a random throw of dart and lands on the outermost ring. It gets a certain amount of points, say x1. It now knows that throwing near that area will yield it an expected value of x1. So, in the next throw, it makes a very slight change of angle and luckily lands in the second outermost right, fetching it x2 points. Since x2 is greater than x1, the robot has achieved a better result and it will learn to throw nearby this area in the future. If the dart had landed even further out than the outermost ring, the robot would keep throwing it near the first throw that it made until it got a better result. 

Over several such trials, the robot keeps learning the better places to throw and makes small detours from those positions until it gets the next better place to throw at. Eventually, it finds the bull's eye and meets the highest points every time.

In the preceding example, your robot is the agent who is trying to throw a dart at the dartboard, which is the environment. Throwing the dart is the action the agent performs on the environment. The points the agent gets act as the reward. The agent, over multiple trials, tries to maximize the reward that it gets by performing the actions.

Some well-known RL algorithms are Monte Carlo, Q-learning, and SARSA.

While we have discussed the three major types of ML, there exists yet another type, which is semi-supervised learning. By the name of the term, you could guess that it would have to do something with a mix of labeled and unlabeled training samples. In most cases, the number of unlabeled training samples exceeds the number of labeled samples.

Semi-supervised learning has been used successfully to produce more efficient results when some labeled samples are added to a problem entirely belonging to unsupervised learning. Also, since only a few samples are labeled, the complexity of supervised learning is avoided. With this approach, we can produce better results than we would get from a purely unsupervised learning system and incur lesser computational cost than a pure supervised learning system. 

We have made ourselves familiar with different types of ML systems. Now, we will learn about some extremely important terminologies related to ML that will help us in the later chapters of this book. 

Any ML system is to be given data. Without data, it is practically impossible to design an ML system. We are not concerned about the quantity of the data as of now, but it is important to keep in mind that we need data to devise an ML system. Once we have that data, we use it for training our ML systems so that they can be used to predict something on the new data (something is a broad term here and it varies from problem to problem). So, the data that is used for training purposes is known as a train set and the data on which the systems are tested is known as a test set. Also, before actually employing the model on the test data, we tend to validate its performance on another set of data, which is called a validation set. Sometimes, we don't get the data in these nice partitions; we just get the data in a raw unfathomable format, which we further process and make these partitions with accordingly. 

Technically, all of the instances in these three different sets are supposed to vary from each other while the distribution in the data is supposed to be the same. Nowadays, many researchers have found critical issues regarding these assumptions and have come up with something called adversarial training, which is out of the scope of this book. 

Bias and variance are very intrinsic to any ML model. Having a good understanding of them really helps in the further assessment of the models. The trade-off between the two is actually used by the practitioners to assess the performance of machine learning systems.

Bias is the set of assumptions that an ML algorithm makes to learn the representations underlying the given data. When the bias is high, it means that the corresponding algorithm is making more assumptions about the data and in the case of low bias, an algorithm makes as little an amount of assumptions as possible. An ML model is said to have a low bias when it performs well on the train set. Some examples of low-bias ML algorithms are k-nearest neighbors and support vector machines while algorithms such as logistic regression and naive Bayes are generally high-bias algorithms. 

Variance in an ML context concerns the information present in the data. Therefore, high variance refers to the quality of how well an ML model has been able to capture the overall information present in the data given to it. Low variance conveys just the opposite. Algorithms such as support vector machines are generally high on variance and algorithms such as naive Bayes are low on variance. 

When an ML model performs very well on the training data but poorly on the data from either the test set or validation set, the phenomenon is referred to as overfitting. There can be several reasons for this; the following are the most common ones: 

• The model is very complex with respect to the data. A decision tree with very high levels and a neural network with many layers are good examples of model complexity in this case. 
• The data has lots of features but very few instances of the population. 

In ML literature, the problem of overfitting is also treated as a problem of high variance. Regularization is the most widely used approach to prevent overfitting.

We have already discussed the concept of bias. A model has a low bias if it performs well on the training data, that is, the model is not making too many assumptions on the data to infer its representation. If the model fails miserably on the training data, it is said that the model has a high bias and the model is underfitting. There can be many reasons for underfitting as well. The following are the most common ones in this case:

• The model is too simple to learn the underlying representation of the data given to it.
• The features of the data have not been engineered well before feeding them to the ML model. The engineering part is very popularly known as feature engineering. 

The discussion of overfitting and underfitting remains incomplete without the following diagram (shown by Andrew Ng during his flagship course, Machine Learning):

The preceding diagram beautifully illustrates underfitting and overfitting in terms of curvea fitting through the data points. It also gives us an idea of a model that generalizes well, that is, performs well on both the train and test sets. The model prediction line in blue is way off the samples, leading to underfitting, while in the case of overfitting, the model captures all points in the training data but does not yield a model that would perform well on data outside training data.

The mistakes that a model makes while predicting during its training phase are collectively referred to as its training error. The mistakes that model makes when tested on either the validation set or the test set are referred to as its generalization error. 

If we were to draw a relationship between these two types of error and bias and variance (and eventually overfitting and underfitting), this would look something like the following (although the relationship may not be linear every time as depicted in the diagrams): 

We will look at a standard ML workflow in the following section.