[IML] early sections update

semester6/iml/main.tex

@@ -15,31 +15,42 @@
 \input{parts/01_regression.tex}
 
 \newpage
-\section{Classification}
+\section{Optimization}
-\input{parts/02_classification.tex}
+\input{parts/02_optimization.tex}
 
 \newpage
+\section{Model Selection}
+\input{parts/03_model_selection.tex}
+
+\newpage
+\section{Regularization}
+\input{parts/04_regularization.tex}
+
+\newpage
+\section{Classification}
+\input{parts/05_classification.tex}
+
 \section{Kernels}
-\input{parts/03_kernels.tex}
+\input{parts/06_kernels.tex}
 
 \newpage
 \section{Neural Networks}
-\input{parts/04_networks.tex}
+\input{parts/07_networks.tex}
 
 \newpage
 \section{Unsupervised Learning}
-\input{parts/05_unsupervised.tex}
+\input{parts/08_unsupervised.tex}
 
 \newpage
 \section{Probabilistic Modelling}
-\input{parts/06_probabilistic.tex}
+\input{parts/09_probabilistic.tex}
 
 \newpage
 \section{Gaussian Mixture Models}
-\input{parts/07_GMM.tex}
+\input{parts/10_GMM.tex}
 
 \newpage
 \section{Language Modeling}
-\input{parts/08_LLM.tex}
+\input{parts/11_LLM.tex}
 
 \end{document}

semester6/iml/parts/01_regression.tex

@@ -32,6 +32,8 @@ $$
     L(f) := \frac{1}{n}\sum_{i=1}^{n} l\bigl( f(x_i), y \bigr) 
 $$
 
+\newpage
+
 \subsection{Multiple Linear Regression}
 
 \textbf{Multiple Linear Regression} directly uses the $x \in \R^d$. \\
@@ -44,9 +46,7 @@ Here, $F_\text{affine} = \bigl\{ f(x) = w^\top x + w_0 \big| w \in \R^d, w_0 \in
     instead search in $F_\text{linear} = \{ f(x) = \hat{w}^\top x | \hat{w} \in \R^{d+1} \}$
 }
 
-\newpage
+\subsection{Loss Functions}
-
-\textbf{Loss Functions}
 
 \definition \textbf{Squared Loss} $\quad l\bigl( f(x),y \bigr) := \bigl( f(x) - y \bigr)^2$\\
 \subtext{Most common Loss Function, but sensitive to outliers.}
@@ -73,9 +73,10 @@ $$
 
 \newpage
 
-\textbf{Linear Regression}
+\subsection{The Normal Equation}
+The normal equation is the basis for the closed form solution of linear regression. (square loss)
 
-To find $\hat{f} := \underset{f \in F_\text{linear}}{\text{arg min}} L(f)$ we just look for $w \in \R^d$.
+To find $\hat{f} := \underset{f \in F_\text{linear}}{\text{arg min}} L(f)$ we look for ideal weights $\hat{w} \in \R^d$.
 $$
     \hat{w} := \underset{w \in \R^d}{\text{arg min}} L(f_w) = \frac{1}{n}\sum_{i=1}^{n}\underbrace{\Bigl( y_i - w^\top x_i \Bigr)^2}_{l\bigl(f(x_i), y_i\bigr)}
 $$
@@ -87,15 +88,74 @@ $$
 $$
 \subtext{The factor $\frac{1}{n}$ is irrelevant for Optimization, it doesn't depend on $w$}
 
-So we find the usual problem:
+So we find a problem familiar from linear algebra:
 $$
     \hat{w} = \underset{w \in \R^d}{\text{arg min}} \bigl\Vert y-Xw \bigr\Vert^2
 $$
 The solution is a stationary point, so:
 $$
-    \nabla_w \bigl\Vert y-Xw \bigr\Vert^2 = 2X^\top(X\hat{w}-y) = 0
+    \nabla_w \bigl\Vert y-Xw \bigr\Vert^2 = 2X^\top(X\hat{w}-y) \overset{!}{=} 0
 $$
-Which yields the \textbf{Normal Equation} from linear algebra.
+Which yields the \textbf{Normal Equation}.
 $$
-    X^\top X\hat{w} = X^\top y
+    \mathbf{X}^\top\mathbf{X}\hat{w} = \mathbf{X}^\top y
 $$
+
+\theorem \textbf{Geometric Interpretation}\\
+$\hat{y} = \mathbf{X}\hat{w}$ for $\hat{w}$ solving $\mathbf{X}^\top\mathbf{X}\hat{w} = \mathbf{X}^\top y$ is the orthogonal projection of $y$ onto $\text{span}(\mathbf{X})$.
+
+\begin{center}
+    \includegraphics[width=0.275\textwidth]{resources/normalEquation.png}\\
+    \subtext{Introduction to Machine Learning (2026), p. 74}
+\end{center}
+
+\newpage
+
+\subsection{Closed Form Solution}
+
+\theorem \textbf{Minimum-Norm Solution}\\
+$$
+    \hat{w} = \Bigl(\mathbf{X}^\top\mathbf{X}\Bigr)^\dagger\mathbf{X}^\top y = \mathbf{X}^\top\Bigl(\mathbf{X}^\top\mathbf{X}\Bigr)^{-1} \mathbf{X}^\top y 
+$$
+$$
+    \text{for} \qquad \hat{w} = \underset{w \in \R^d}{\text{arg min}} \bigl\Vert w \bigr\Vert^2
+$$
+
+{\footnotesize
+    \remark The computational cost for this is $\mathcal O (nd^2+d^3)$.
+}
+
+The closed form solution depends on $\text{rank}(\mathbf{X})$.
+
+Assuming $d \leq n$ and $\text{rank}(\mathbf{X}) = d$: $(\mathbf{X}^\top\mathbf{X})^{-1}$ exists.
+$$
+    \hat{w} = (\mathbf{X}^\top\mathbf{X})^{-1}\mathbf{X}^\top y \qquad (\text{unique})
+$$
+Assuming $d > n$ or $\text{rank}(\mathbf{X}) < d$ we have $|\ker(\mathbf{X})|=\infty$ and there are infinite solutions. The pseudo-inverse provides the minimum-norm solution:
+$$
+    \hat{w} = \Bigl(\mathbf{X}^\top\mathbf{X}\Bigr)^\dagger \mathbf{X}^\top y
+$$
+{\footnotesize
+    \remark If $\text{rank}(\mathbf{X})=d$, then $\bigl(\mathbf{X}^\top\mathbf{X}\bigr)^\dagger = \bigl(\mathbf{X}^\top\mathbf{X}\bigr)^{-1}$.
+}
+
+\subsection{Non-Linear Least Squares}
+
+To expand Linear Regression to Non-linear functions, feature maps are used on $x$: $\phi: \R^d \to \R^p$.
+$$
+    f_w(x) = \sum_{j=1}^{p} w_j^\top \phi_j(x)
+$$
+This induces a function class different from $F_\text{linear}$:
+$$
+    F_\phi = \biggl\{ f_w(x) = \sum_{j=1}^{p} w_j^\top \phi_j(x) \ \bigg|\ w \in \R^p \biggr\}
+$$
+But the optimization problem remains the same:\\
+\subtext{$\Phi \in \R^{n \times p}$ now replaces $\mathbf{X} \in \R^{n \times d}$.}
+$$
+    \hat{w} = \underset{w\in\R^p}{\text{arg min}} \Bigl\Vert y - \Phi w \Bigr\Vert^2
+$$
+
+\begin{center}
+    \includegraphics[width=0.2\textwidth]{resources/nonlinearLeastSquares.png}\\
+    \subtext{Introduction to Machine Learning (2026), p. 79}
+\end{center}

semester6/iml/parts/02_optimization.tex

@@ -0,0 +1,54 @@
+\textbf{Problem}: Finding $\hat{w}$ for $l$ with no closed form solution.\\
+\subtext{or if the closed form solution is too expensive to compute.}
+
+\textbf{Solution}: Iterative optimization methods.
+
+\begin{algorithm}
+    \caption{Iterative Optimization}
+    $t \gets 0$ \;
+    $w^{(0)} \gets w_\text{initial}$ \;
+    \SetKwRepeat{Do}{repeat}{until}
+        \Do{\text{Stopping Criterion}}{ 
+            $w^{(t+1)} \gets w^{(t)} + \tilde{\eta}_t v^{(t)}$ \;
+            $t \gets t+1$
+    }
+    \textbf{return} $w^{(t)}$
+\end{algorithm}
+
+{\footnotesize
+    \notation The update takes the form $\tilde{\eta}_t v^{(t)}$. $v^{(t)}$ is the update direction, $\tilde{\eta}_t$ is the step size.
+}
+
+\subsection{Gradient Descent}
+
+Intuitively: go in the direction $v^{(t)}$ where $L$ decreases most.
+
+\lemma $-\nabla L(w^{(t)})$ is the direction of steepest descent.\\
+\subtext{Assuming diff.-able $L$. Provable via Taylor expansion \& Cauchy-Schwarz.}
+
+\definition \textbf{Gradient Descent Update Step}
+\begin{align*}
+        w^{(t+1)} &= w^{(t)} - \tilde{\eta}_t\cdot \frac{\nabla L(w^{(t)})}{\Vert \nabla L(w^{(t)}) \Vert}   & (\text{Normalized}) \\
+        w^{(t+1)} &= w^{(t)} - \eta\cdot \nabla L(w^{(t)})                                         & (\text{Unnormalized})
+\end{align*}
+
+Unnormalized gradient descent takes advantage of $\Vert \nabla L(w^{(t)}) \Vert$:
+{\small
+    \begin{itemize}
+        \item $\Vert \nabla L(w^{(t)}) \Vert$ small $\mapsto$ close to stat. point $\mapsto$ small steps.
+        \item $\Vert \nabla L(w^{(t)}) \Vert$ large $\mapsto$ far from stat. point $\mapsto$ large steps.  
+    \end{itemize}
+}
+Stopping criterion uses the same idea: $\Vert w^{t} - w^{t+1} \Vert < \epsilon$ or equivalently $\Vert \nabla L(w^{(t)}) \Vert < \epsilon$.
+
+\begin{algorithm}
+    \caption{Gradient Descent}
+    $t \gets 0$ \;
+    $w^{(0)} \gets w_\text{initial}$ \;
+    \SetKwRepeat{Do}{repeat}{until}
+        \Do{\text{$\Vert w^{t} - w^{t+1} \Vert < \epsilon$}}{ 
+            $w^{(t+1)} \gets w^{(t)} - \eta \nabla L(w^{(t)})$ \;
+            $t \gets t+1$
+    }
+    \textbf{return} $w^{(t)}$
+\end{algorithm}