There are different approaches to create and train ML models. In this section, we show what these approaches are and how they differ. Apart from the approach we use to create a ML model, there are also parameters that manage how this model behaves in the training and evaluation processes. Model parameters can be divided into two distinct groups, which should be configured in different ways. The last crucial part of the ML process is a technique that we use to train a model. Usually, the training technique uses some numerical optimization algorithm that finds the minimal value of a target function. In ML, the target function is usually called a loss function and is used for penalizing the training algorithm when it makes errors. We discuss these concepts more precisely in the following sections.

We can divide ML approaches into two techniques, as follows:

• Supervised learning is an approach based on the use of labeled data. Labeled data is a set of known data samples with corresponding known target outputs. Such a kind of data is used to build a model that can predict future outputs.
• Unsupervised learning is an approach that does not require labeled data and can search hidden patterns and structures in an arbitrary kind of data.

Let's have a look at each of the techniques in detail.

Supervised ML algorithms usually take a limited set of labeled data and build models that can make reasonable predictions for new data. We can split supervised learning algorithms into two main parts, classification and regression techniques, described as follows:

• Classification models predict some finite and distinct types of categories—this could be a label that identifies if an email is spam or not, or whether an image contains a human face or not. Classification models are applied in speech and text recognition, object identification on images, credit scoring, and others. Typical algorithms for creating classification models are Support Vector Machine (SVM), decision tree approaches, k-nearest neighbors (KNN), logistic regression, Naive Bayes, and neural networks. The following chapters describe the details of some of these algorithms.
• Regression models predict continuous responses such as changes in temperature or values of currency exchange rates. Regression models are applied in algorithmic trading, forecasting of electricity load, revenue prediction, and others. Creating a regression model usually makes sense if the output of the given labeled data is real numbers. Typical algorithms for creating regression models are linear and multivariate regressions, polynomial regression models, and stepwise regressions. We can use decision tree techniques and neural networks to create regression models too. The following chapters describe the details of some of these algorithms.

Unsupervised learning algorithms do not use labeled datasets. They create models that use intrinsic relations in data to find hidden patterns that they can use for making predictions. The most well-known unsupervised learning technique is clustering. Clustering involves dividing a given set of data in a limited number of groups according to some intrinsic properties of data items. Clustering is applied in market researches, different types of exploratory analysis, deoxyribonucleic acid (DNA) analysis, image segmentation, and object detection. Typical algorithms for creating models for performing clustering are k-means, k-medoids, Gaussian mixture models, hierarchical clustering, and hidden Markov models. Some of these algorithms are explained in the following chapters of this book.

We can interpret ML models as functions that take different types of parameters. Such functions provide outputs for given inputs based on the values of these parameters. Developers can configure the behavior of ML models for solving problems by adjusting model parameters. Training a ML model can usually be treated as a process of searching the best combination of its parameters. We can split the ML model's parameters into two types. The first type consists of parameters internal to the model, and we can estimate their values from the training (input) data. The second type consists of parameters external to the model, and we cannot estimate their values from training data. Parameters that are external to the model are usually called hyperparameters.

Internal parameters have the following characteristics:

• They are necessary for making predictions.
• They define the quality of the model on the given problem.
• We can learn them from training data.
• Usually, they are a part of the model.

If the model contains a fixed number of internal parameters, it is called parametric. Otherwise, we can classify it as non-parametric.

Examples of internal parameters are as follows:

• Weights of artificial neural networks (ANNs)
• Support vector values for SVM models
• Polynomial coefficients for linear regression or logistic regression

On the other hand, hyperparameters have the following characteristics:

• They are used to configure algorithms that estimate model parameters.
• The practitioner usually specifies them.
• Their estimation is often based on using heuristics.
• They are specific to a concrete modeling problem.

It is hard to know the best values for a model's hyperparameters for a specific problem. Also, practitioners usually need to perform additional research on how to tune required hyperparameters so that a model or a training algorithm behaves in the best way. Practitioners use rules of thumb, copying values from similar projects, as well as special techniques such as grid search for hyperparameter estimation.

Examples of hyperparameters are as follows:

• C and sigma parameters used in the SVM algorithm for a classification quality configuration
• The learning rate parameter that is used in the neural network training process to configure algorithm convergence
• The k value that is used in the KNN algorithm to configure the number of neighbors

Model parameter estimation usually uses some optimization algorithm. The speed and quality of the resulting model can significantly depend on the optimization algorithm chosen. Research on optimization algorithms is a popular topic in industry, as well as in academia. ML often uses optimization techniques and algorithms based on the optimization of a loss function. A function that evaluates how well a model predicts on the data is called a loss function. If predictions are very different from the target outputs, the loss function will return a value that can be interpreted as a bad one, usually a large number. In such a way, the loss function penalizes an optimization algorithm when it moves in the wrong direction. So, the general idea is to minimize the value of the loss function to reduce penalties. There is no one universal loss function for optimization algorithms. Different factors determine how to choose a loss function. Examples of such factors are as follows:

• Specifics of the given problem—for example, if it is a regression or a classification model
• Ease of calculating derivatives
• Percentage of outliers in the dataset

In ML, the term optimizer is used to define an algorithm that connects a loss function and a technique for updating model parameters in response to the values of the loss function. So, optimizers tune ML models to predict target values for new data in the most accurate way by fitting model parameters. There are many optimizers: Gradient Descent, Adagrad, RMSProp, Adam, and others. Moreover, developing new optimizers is an active area of research. For example, there is the ML and Optimization research group at Microsoft (located in Redmond) whose research areas include combinatorial optimization, convex and non-convex optimization, and their application in ML and AI. Other companies in the industry also have similar research groups; there are many publications from Facebook Research, Amazon Research, and OpenAI groups.