The concept of **machine learning** it is closely linked with that of neural networks and, in general, with artificial intelligence applications. This is an area that deals with the development of algorithms and models that allow the machine to learn from past data, identify *pattern*make predictions and possibly make decisions without being explicitly programmed for specific tasks.

Il machine learning (**machine learning**, in Europen) can be divided into two main categories: supervised and unsupervised. In the’**supervised learning**the model is trained on a dataset of known inputs and outputs, while in learning **unsupervised** the model must identify *pattern* and structures in the data without the help of *output* known and approved.

## The free Amazon MLU-Explain course to learn machine learning

Amazon, as part of its educational project *Machine Learning University* (MLU), offers a **free course** called MLU-Explain. This is an initiative designed to teach machine learning theory and its practical application in an accessible and effective way.

Main objective of **MLU-Explain** is to teach crucial machine learning concepts through a series of visual steps and interactive descriptions. The first module focused on **neural networks** explores essential components for many of the most popular and highly regarded algorithms: ChatGPT, Stable Diffusion and many others.

We then move on to delve deeper into the concept of *equality of odds* (equality of probabilities): the course signed by Amazon gives ample space to the metrics used to evaluate and reduce the so-called to a minimum ** bias**. The latter represent a systematic deviation or distortion in the collection, analysis and interpretation of data, leading to an alteration of conclusions or predictions. THE

*bias*they can manifest themselves in various ways, negatively influencing the

**correctness**and the

**reliability**of a system, algorithm or model.

I *bias* in generative models can unintentionally harm or create negative stereotypes against underrepresented or historically disadvantaged groups of subjects. The concept of *equality of odds *mira a **smooth out the error** that a model commits when it predicts categorical outcomes for different groups.

### Logistic regression and reinforcement learning

An interactive module illustrates how logistic regression can be applied to **binary classification**. Using several examples, experts show how a model can learn from data and make accurate predictions in a practical context. There **logistic regression** is a supervised learning algorithm used to classify data into categories or classes, predicting the probability that an observation belongs to a particular class based on its specific characteristics.

Although the *logistic regression* can be extended to more than two categories, it is often used for binary classification, i.e. to predict which of the two groups a data item belongs to or whether an event will occur or not.

The course also explores the so-called **reinforcement learning**focusing on the dilemma *exploration-exposure*. Through interactive scenarios, it explains how agents can learn to make optimal decisions in complex environments, balancing the discovery of new behaviors and the exploitation of already known ones.

Reinforcement learning is a branch of machine learning that focuses on how “intelligent agents” can learn **interacting with the environment** surrounding. Unlike supervised and unsupervised learning, in reinforcement learning the model learns from **consequences of his actions** without knowing a priori the right operation to perform.

### ROC curve and AUC: performance evaluation of classifiers

A detailed analysis of the *ROC curve* and of the so-called *Area Under the Curve* (AUC) offer in-depth insight into how to evaluate the performance of a **classifier**.

A “classifier” is a machine learning model trained to assign objects or data to one or more classes, based on their intrinsic characteristics. In other words, a classifier takes as input the characteristics of an instance and provides as output the label of the class to which that instance belongs. The “performance” of a classifier refers to its ability to make accurate predictions.

The course also introduces the **cross-validation technique** *K-Fold* as a tool to improve error estimates. Furthermore, by explaining the importance of dividing data into training, test and validation sets, the course provides the foundation for ensuring that **models are accurate** and behave well even when they find themselves examining new data, never seen before.

### Decision trees, balance between bias and bariance, double descent phenomenon

The MLU-Explain course concludes by examining three key concepts in the field of machine learning. The **decision trees**, for example, represent an intuitive approach in machine learning, used for both classification and regression problems. These algorithms work by creating a decision tree structure based on the characteristics of the training data. Each node in the tree represents a question about a specific characteristic, while the branches correspond to the possible answers. This structure allows a visual understanding of the decision-making process and facilitates the interpretation of the model.

Il **balance between bias and variance** is a crucial consideration in the model development process. In fact, the bias can also be defined as the error introduced by the simplifications made by the model, while the variance measures how sensitive the model is to variations in the training data. Finding the right compromise between these two components is essential to avoid the occurrence of phenomena capable of negatively influencing the performance of the models.

The concept of **double descent** has emerged in recent years as one of the fundamental aspects to take into consideration when working on machine learning solutions. Contrary to the traditional belief that a model’s error decreases monotonically with complexity, double descent highlights a “U” curve. In other words, after an initial phase of **error descent**increasing the complexity of the model presents a second phase of error reduction.

*Opening image credit: iStock.com – MF3d*