diff --git a/semester6/iml/main.pdf b/semester6/iml/main.pdf
index 74037e9..ee25a8f 100644
Binary files a/semester6/iml/main.pdf and b/semester6/iml/main.pdf differ
diff --git a/semester6/iml/parts/04_networks.tex b/semester6/iml/parts/04_networks.tex
index 3d3d59d..fcf0f3b 100644
--- a/semester6/iml/parts/04_networks.tex
+++ b/semester6/iml/parts/04_networks.tex
@@ -11,6 +11,12 @@ $$
 Where $\phi(x,\Theta) = \Bigl( \phi_1(x;\theta_1),\ldots,\phi_m(x;\theta_m) \Bigr)$.\\
 \subtext{$\theta_i$ is the $i$th row of $\Theta$, i.e. $\theta_i := (\Theta)_{i,:}$}
 
+More compact, in terms of $\Theta$, which combines $w, \phi$:
+$$
+    \Theta^* := \underset{\Theta}{\text{arg min}}\Bigl( L(\Theta; \mathcal{D}) \Bigr) = \underset{\Theta}{\text{arg min}} \Biggl( \frac{1}{n} \sum_{i=1}^{n} l\Bigl( \Theta;x_i,y_i \Bigr) \Biggr)
+$$
+\subtext{$\Theta$ may also encapsulate $w,\phi$ for multiple layers, depending on definition}
+
 \subsection{Definitions}
 
 \definition \textbf{Activation Function}\\
@@ -94,10 +100,19 @@ These are called \textit{fully connected}, since every node in a layer is connec
     \subtext{\textit{Introduction to Machine Learning (2026), p. 183}}
 \end{center}
 
-\definition \textbf{Forward Propagation}
+\notation Weights: $\textbf{W}^{(i)} := \Bigl[ w_{k,l}^{(i)} \Bigr]$, Biases: $b^{(i)}_k$\\
+and $\Theta = \Bigl(\textbf{W}^{(1)},\ldots,\textbf{W}^{(L)}, b^{(1)},\ldots,b^{(L)}\Bigr)$ (All parameters)\\
+\subtext{$w_{k,l}^{(i)}$: "Weight at layer $i$ to node $k$ from node $l$"} 
+
+\newpage
+\subsection{Forward Propagation}
+
+
+How can we make predictions, i.e. how can $\hat{f}$ be evaluated?
+
+\definition \textbf{Forward Propagation}\\
+\subtext{This is just the computation for $1$-layer ANN generalized for $L$ layers}
 
-How can $\hat{f}$ be evaluated?\\
-\subtext{This is just the computation for $1$-layer ANN generalized}
 
 \begin{algorithm}
     \caption{Forward Propagation}
@@ -108,4 +123,108 @@ How can $\hat{f}$ be evaluated?\\
     }
     $f \gets \textbf{W}^{(L)}h^{(L-1)}+b^{(L)}$ \\
     \Return f
-\end{algorithm}
\ No newline at end of file
+\end{algorithm}
+
+\subsection{Backwards Propagation}
+
+How can we get all gradients needed for model training?
+
+\definition \textbf{Backwards Propagation}
+
+\textbf{Intuition}: An efficient way to get the gradients is to reuse results from forward prop. and previous steps. This works best when starting at the back, at $\nabla_{\textbf{W}^{(L)}}l$.
+
+\textbf{Goal}: $\nabla_{\textbf{W}^{(1)}}l,\ldots,\nabla_{\textbf{W}^{(L)}} l, \nabla_{b^{(1)}}l,\ldots,\nabla_{b^{(L)}}l$
+
+\textbf{Step 1}: Calculate $\nabla_{\textbf{W}^{(L)}}l$, i.e. start from the back.
+\begin{align*}
+    \nabla_{\textbf{W}^{(L)}}l  &= \frac{\partial l}{\partial \textbf{W}^{L}} \\
+                                &= \frac{\partial l}{\partial f}\cdot\frac{\partial f}{\partial \textbf{W}^{(L)}} & \text{(Chain Rule)} \\
+                                &= \frac{\partial l}{\partial f}\cdot\begin{bmatrix}
+                                    \bigl( h^{(L-1)} \bigr)^\top    \\
+                                    \vdots                          \\
+                                    \bigl( h^{(L-1)} \bigr)^\top    
+                                \end{bmatrix}                                                                     & (f = \textbf{W}^{(L)}h^{(L-1)} + b^{(L)}) \\
+                                &= \nabla_f l \cdot\begin{bmatrix}
+                                    \bigl( h^{(L-1)} \bigr)^\top    \\
+                                    \vdots                          \\
+                                    \bigl( h^{(L-1)} \bigr)^\top    
+                                \end{bmatrix}                                                                     & \Biggl(\frac{\partial l}{\partial f} = \nabla_f l\Biggr)
+\end{align*}
+Notice how $h^{(L-1)}$ was computed during forward prop.
+
+\newpage
+
+\textbf{Step 2}: Calculate $\nabla_{\textbf{W}^{(L-1)}}l$.
+\begin{align*}
+    \nabla_{\textbf{W}^{(L-1)}}l    &= \underbrace{\frac{\partial l}{\partial f}}_{\text{(1)}}\cdot\underbrace{\frac{\partial f}{\partial h^{(L-1)}}}_{\text{(2)}}\cdot\underbrace{\frac{\partial h^{(L-1)}}{\partial z^{(L-1)}}}_\text{(3)}\cdot\underbrace{\frac{z^{(L-1)}}{\partial \textbf{W}^{(L-1)}}}_\text{(4)}      & \text{(Chain Rule)} \\
+\end{align*}
+\begin{enumerate}
+    \item Already done in Step 1.
+    \item Already done in forward propagation, equal to $\textbf{W}^{(L)}$:
+    $$
+        f \overset{\text{def}}{=} \textbf{W}^{(L)}h^{(L-1)}+b^{(L)} \implies \frac{\partial f}{\partial h^{(L-1)}} = \textbf{W}^{(L)}
+    $$
+    \item \textbf{Not done.} Needs to be calculated:
+    \begin{align*}
+        \frac{\partial h^{(L-1)}}{\partial z^{(L-1)}}   &= \frac{\partial \psi\bigl( z^{(L-1)} \bigr)}{\partial z^{(L-1)}}  \\
+                                                        &= \text{diag}\Bigl( \psi'\bigl( z^{(L-1)} \bigr) \Bigr)            \\
+                                                        &= \begin{bmatrix}
+                                                            \psi'\Bigl(z_1^{(L-1)}\Bigr) & 0 & \cdots & 0 \\
+                                                            0 & \psi'\Bigl(z_2^{(L-1)}\Bigr) & \cdots & 0 \\
+                                                            \vdots & \vdots & \ddots & \vdots             \\
+                                                            0 & 0 & \cdots & \psi\Bigl(z_n^{(L-1)}\Bigr)  \\
+                                                        \end{bmatrix}  
+    \end{align*}
+    \item Already done in forward propagation, analogous to step 1. 
+    $$
+        \frac{\partial z^{(L-1)}}{\partial \textbf{W}^{(L-1)}} =
+        \begin{bmatrix}
+            \bigl( h^{(L-2)} \bigr)^\top    \\
+            \vdots                          \\
+            \bigl( h^{(L-2)} \bigr)^\top
+        \end{bmatrix}
+    $$
+\end{enumerate}
+
+\textbf{Step $i \leq L$}: Calculate $\nabla_{\textbf{W}^{(L-i)}}l$ Analogoues to step 2.\\
+\subtext{The biases $\nabla_{b^{(l)}}l$ are analogous.}
+
+\newpage
+
+\subsection{Optimization}
+
+\textbf{Problem}: How can we train the model, i.e find $\Theta^*$?
+$$
+    \Theta^* := \underset{\Theta}{\text{arg min}}\Bigl( L(\Theta; \mathcal{D}) \Bigr) = \underset{\Theta}{\text{arg min}} \Biggl( \frac{1}{n} \sum_{i=1}^{n} l\Bigl( \Theta;x_i,y_i \Bigr) \Biggr)
+$$
+
+{\footnotesize
+    \remark $L(\Theta;\mathcal{D})$ is generally not convex.\\
+    {\color{gray}
+        i.e. local minima, saddle points may exist
+    }
+
+    \remark $\dim(\Theta)$ is the total param. count of NN, may be very large
+}
+
+\textbf{Solution}: Gradient Descent (with optimizations)
+\begin{itemize}
+    \item Stochastic Gradient Descent\\
+    \subtext{(Why? $\dim(\Theta)$ is very large, $\nabla_\Theta l(\Theta;x_i,y_i)$ are expensive)}
+    \item Minibatch Gradient Descent\\
+    \subtext{(Why? $\mathcal{D}$ may be very large, so there are \textit{many} gradients)}
+\end{itemize}
+
+The standard GD update for $\Theta$ is:
+$$
+    \Theta^{t+1} = \Theta^t - \eta_t\cdot\nabla_\Theta L\Bigl( \Theta;\mathcal{D} \Bigr)
+$$
+In Minibatch GD, this becomes:\\
+\subtext{Where $\mathcal{S} \subset \{1,\ldots,n\}$}
+$$
+    \Theta^{t+1} = \Theta^t - \eta_t\cdot\nabla_\Theta L\Biggl( \frac{1}{|\mathcal{S}|}\sum_{i\in \mathcal{S}} l\Bigl( \Theta^t; x_i,y_i \Bigr) \Biggr)
+$$
+
+{\footnotesize
+    \remark An advantage: If $\Theta^t$ approaches a stationary point (which isn't the global minimu), GD will converge, but MB-GD may not converge.
+}
\ No newline at end of file