Files
unibo-ai-notes/src/statistical-and-mathematical-methods-for-ai/sections/_machine_learning.tex
2023-10-12 21:22:57 +02:00

84 lines
4.2 KiB
TeX

\chapter{Machine learning}
\section{Models}
\begin{description}
\item[Function model] \marginnote{Function model}
The model (predictor) is a deterministic function:
\[ f: \mathbb{R}^D \rightarrow \mathbb{R} \]
In this course, only linear functions are considered:
\[ f_\vec{\uptheta}(\vec{x}) = \uptheta_0 + \uptheta_1 x_1 + \dots + \uptheta_D x_D = \vec{\uptheta}^T \vec{x} \]
where $\vec{x} = \begin{pmatrix} 1, x_1, \dots, x_D \end{pmatrix}$ is the input vector and
$\vec{\uptheta} = \begin{pmatrix} \uptheta_0, \dots, \uptheta_D \end{pmatrix}$ is the parameter vector.
\item[Probabilistic model] \marginnote{Probabilistic model}
The model is a multivariate probabilistic distribution.
\end{description}
\section{Learning}
\subsection{Empirical risk minimization}
\marginnote{Empirical risk minimization}
Used for function models.
Let $(\vec{x}_n, y_n)$ be a dataset of $N$ elements
where $\vec{x}_n \in \mathbb{R}^D$ are the examples and $y_n \in \mathbb{R}$ are the labels.
We want to estimate a predictor $f_\vec{\uptheta}(\vec{x}) = \vec{\uptheta}^T \vec{x}$ with parameters $\vec{\uptheta}$
such that, with the ideal parameters $\vec{\uptheta}^*$, it fits the data well:
\[ f_{\vec{\uptheta}^*}(\vec{x}_n) \approx y_n \]
We denote the output of the estimator as $\hat{y}_n = f_\vec{\uptheta}(\vec{x}_n)$.
\begin{description}
\item[Loss function] \marginnote{Loss function}
A loss function $\ell(y_n, \hat{y}_n)$ indicates how a predictor fits the data.
An assumption commonly made in machine learning is that
the dataset $(\vec{x}_n, y_n)$ is independent and identically distributed.
Therefore, the empirical mean is a good estimate of the population mean.
\begin{description}
\item[Empirical risk] \marginnote{Empirical risk}
Given the example matrix $\matr{X} = \begin{pmatrix} \vec{x}_1, \dots, \vec{x}_N \end{pmatrix} \in \mathbb{R}^{N \times D}$
and the label vector $\vec{y} = \begin{pmatrix} y_1, \dots, y_N \end{pmatrix} \in \mathbb{R}^N$.
The empirical risk is given by the average loss:
\[ \textbf{R}_\text{emp}(f_\vec{\uptheta}, \matr{X}, \vec{y}) = \frac{1}{N} \sum_{n=1}^{N} \ell(y_n, \hat{y}_n) \]
\begin{example}[Least-squares loss] \marginnote{Least-squares loss}
The least-squares loss is defined as:
\[ \ell(y_n, \hat{y}_n) = (y_n - \hat{y}_n)^2 \]
Therefore, the minimization task is:
\[
\min_{\vec{\uptheta} \in \mathbb{R}^D} \frac{1}{N} \sum_{n=1}^{N} (y_n - f_\vec{\uptheta}(\vec{x}_n))^2 =
\min_{\vec{\uptheta} \in \mathbb{R}^D} \frac{1}{N} \sum_{n=1}^{N} (y_n - \vec{\uptheta}^T\vec{x}_n)^2 =
\min_{\vec{\uptheta} \in \mathbb{R}^D} \frac{1}{N} \Vert \vec{y} - \matr{X}\vec{\uptheta} \Vert^2
\]
\end{example}
\item[Expected risk] \marginnote{Expected risk}
The expected risk is defined as:
\[ \textbf{R}_\text{true}(f_\vec{\uptheta}) = \mathbb{E}_{\vec{x}, y}[\ell(y, f_\vec{\uptheta}(\vec{x}_\text{test}))] \]
where the parameters $\vec{\uptheta}$ are fixed and the samples are taken from a test set.
\item[Overfitting] \marginnote{Overfitting}
A predictor $f_\vec{\uptheta}$ is overfitting when $\textbf{R}_\text{emp}(f, \matr{X}_\text{train}, \vec{y}_\text{train})$
underestimates $\textbf{R}_\text{true}(f_\vec{\uptheta})$ (i.e. the loss on the training set is low, but on the test set is high).
\item[Regularization] \marginnote{Regularization}
Method that introduces a penalty term to the loss that
helps to find a compromise between the accuracy and the complexity of the solution:
\[ \bar{\ell}(y_n, \hat{y}_n) = \ell(y_n, \hat{y}_n) + \lambda \mathcal{R}(\vec{\uptheta}) \]
where $\lambda \in \mathbb{R}^+$ is the regularization parameter and $\mathcal{R}$ is the penalty.
\end{description}
\end{description}
\subsection{Maximum likelihood}
\marginnote{Maximum likelihood}
Used for probabilistic models.