Masterarbeit/TeX/introduction_nn.tex

\section{Introduction to Neural Networks}

Neural Networks (NN) are a mathematical construct inspired by the
connection of neurons in nature. It consists of an input and output
layer with an arbitrary amount of hidden layers between them. Each
layer consits of a numer of neurons (nodes) with the number of nodes
in the in-/output layers corresponding to the dimensions of the
in-/output.\par
Each neuron recieves the output of all layers in the previous layers,
except for the input layer, which recieves the components of the input.

\tikzset{%
  every neuron/.style={
    circle,
    draw,
    minimum size=1cm
  },
  neuron missing/.style={
    draw=none, 
    scale=1.5,
    text height=0.333cm,
    execute at begin node=\color{black}$\vdots$
  },
}
\begin{figure}[h!]
  \center
  
  % \fbox{
  
  \resizebox{\textwidth}{!}{%
    \begin{tikzpicture}[x=1.75cm, y=1.75cm, >=stealth]
      \tikzset{myptr/.style={decoration={markings,mark=at position 1 with %
            {\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate}}}
      
      \foreach \m/\l [count=\y] in {1,2,3,missing,4}
      \node [every neuron/.try, neuron \m/.try] (input-\m) at (0,2.5-\y) {};
      
      \foreach \m [count=\y] in {1,missing,2}
      \node [every neuron/.try, neuron \m/.try ] (hidden1-\m) at (2,2-\y*1.25) {};
      
      \foreach \m [count=\y] in {1,missing,2}
      \node [every neuron/.try, neuron \m/.try ] (hidden2-\m) at (5,2-\y*1.25) {};
        
      \foreach \m [count=\y] in {1,missing,2}
      \node [every neuron/.try, neuron \m/.try ] (output-\m) at (7,1.5-\y) {};
      
      \foreach \l [count=\i] in {1,2,3,d_i}
      \draw [myptr] (input-\i)+(-1,0) -- (input-\i)
      node [above, midway] {$x_{\l}$};
      
      \foreach \l [count=\i] in {1,n_1}
      \node [above] at (hidden1-\i.north) {$\mathcal{N}_{1,\l}$};
      
      \foreach \l [count=\i] in {1,n_l}
      \node [above] at (hidden2-\i.north) {$\mathcal{N}_{l,\l}$};
      
      \foreach \l [count=\i] in {1,d_o}
      \draw [myptr] (output-\i) -- ++(1,0)
      node [above, midway] {$O_{\l}$};
      
      \foreach \i in {1,...,4}
      \foreach \j in {1,...,2}
      \draw [myptr] (input-\i) -- (hidden1-\j);
      
      \foreach \i in {1,...,2}
      \foreach \j in {1,...,2}
      \draw [myptr] (hidden1-\i) -- (hidden2-\j);
      
      \foreach \i in {1,...,2}
      \foreach \j in {1,...,2}
      \draw [myptr] (hidden2-\i) -- (output-\j);
      
      \node [align=center, above] at (0,2) {Input\\layer};
      \node [align=center, above] at (2,2) {Hidden \\layer $1$};
      \node [align=center, above] at (5,2) {Hidden \\layer $l$};
      \node [align=center, above] at (7,2) {Output \\layer};
      
      \node[fill=white,scale=1.5,inner xsep=10pt,inner ysep=10mm] at ($(hidden1-1)!.5!(hidden2-2)$) {$\dots$};
        
    \end{tikzpicture}}%}
  \caption{Illustration of a neural network with $d_i$ inputs, $l$
    hidden layers with $n_{\cdot}$ nodes in each layer, as well as
    $d_o$ outputs.
  }
\end{figure}

\subsection{Nonlinearity of Neural Networks}



\begin{figure}
  \begin{tikzpicture}[x=1.5cm, y=1.5cm, >=stealth]

    
    \tikzset{myptr/.style={decoration={markings,mark=at position 1 with %
      {\arrow[scale=1.5,>=stealth]{>}}},postaction={decorate}}}
    
    \node [circle, draw, fill=black, inner sep = 0pt, minimum size =
    1.5mm, left] (i_1) at (0, 2.5) {};
    \node [align=left, left] at (-0.125, 2.5) {\(i_1\)};
    \node [circle, draw, fill=black, inner sep = 0pt, minimum size =
    1.5mm] (i_2) at (0, 1.25) {};
    \node [align=left, left] at (-0.125, 1.25) {\(i_2\)};
    \node [neuron missing] (i_3) at (0, 0) {};
    \node [circle, draw, fill=black, inner sep = 0pt, minimum size =
    1.5mm] (i_4) at (0, -1.25) {};
    \node [align=left, left] at (-0.125, -1.25) {\(i_m\)};
    \draw[decoration={calligraphic brace,amplitude=5pt, mirror}, decorate, line width=1.25pt]
    (-0.6,2.7) -- (-0.6,-1.45) node [black, midway, xshift=-0.6cm, left] {Inputs};

    \node [align = center, above] at (1.25, 3) {Synaptic\\weights};
    \node [every neuron] (w_1) at (1.25, 2.5) {\(w_{k, 1}\)};
    \node [every neuron] (w_2) at (1.25, 1.25) {\(w_{k, 2}\)};
    \node [neuron missing] (w_3) at (1.25, 0) {};
    \node [every neuron] (w_4) at (1.25, -1.25) {\(w_{k, m}\)};

    \node [circle, draw] (sig) at (3, 0.625) {\Large\(\sum\)};
    \node [align = center, below] at (3, 0) {Summing \\junction};

    \node [draw, minimum size = 1.25cm] (act) at (4.5, 0.625)
    {\(\sigma(.)\)};
    \node [align = center, above] at (4.5, 1.25) {Activation \\function}; 

    \node [circle, draw, fill=black, inner sep = 0pt, minimum size =
    1.5mm] (b) at (3, 2.5) {};
    \node [align = center, above] at (3, 2.75) {Bias \\\(b_k\)};

    \node [align = center] (out) at (6, 0.625) {Output \\\(o_k\)};


    \draw [myptr] (i_1) -- (w_1);
    \draw [myptr] (i_2) -- (w_2);
    \draw [myptr] (i_4) -- (w_4);

    \draw [myptr] (w_1) -- (sig);
    \draw [myptr] (w_2) -- (sig);
    \draw [myptr] (w_4) -- (sig);

    \draw [myptr] (b) -- (sig);

    \draw [myptr] (sig) -- (act);

    \draw [myptr] (act) -- (out);

    % \foreach \m [count=\y] in {1,2,missing,3,4}
    % \node [every neuron/.try, neuron \m/.try ] (hidden-\m) at (1.25,3.25-\y*1.25) {\(w_{k,\y}\)};

    % \foreach \m [count=\y] in {1}
    % \node [every neuron/.try, neuron \m/.try ] (output-\m) at (2.5,0.5-\y) {};

    % \foreach \l [count=\i] in {1}
    % \draw [<-] (input-\i) -- ++(-1,0)
    % node [above, midway] {$x$};

    % \foreach \l [count=\i] in {1,2,n-1,n}
    % \node [above] at (hidden-\i.north) {$\mathcal{N}_{\l}$};

    % \foreach \l [count=\i] in {1,n_l}
    % \node [above] at (output-\i.north) {};

    % \foreach \l [count=\i] in {1}
    % \draw [->] (output-\i) -- ++(1,0)
    % node [above, midway] {$y$};

    % \foreach \i in {1}
    % \foreach \j in {1,2,...,3,4}
    % \draw [->] (input-\i) -- (hidden-\j);

    % \foreach \i in {1,2,...,3,4}
    % \foreach \j in {1}
    % \draw [->] (hidden-\i) -- (output-\j);

  \end{tikzpicture}
  \caption{Structure of a single neuron}
\end{figure}

\clearpage
\subsection{Training Neural Networks}

After a neural network model is designed, like most statistical models
it has to be fit to the data. In the machine learning context this is
often called ``training'' as due to the complexity and amount of
variables in these models they are fitted iteratively to the data,
``learing'' the properties of the data better with each iteration.

There are two main categories of machine learning models, being
supervised and unsupervised learners. Unsupervised learners learn
structure in the data without guidance form outside (as labeling data
beforehand for training) popular examples of this are clustering
algorithms\todo{quelle}. Supervised learners on the other hand are as
the name suggest supervised during learning. This generally amounts to
using data with the expected response (label) attached to each
data-point in fitting the model, where usually some distance between
the model output and the labels is minimized.

\subsubsection{Interpreting the Output / Classification vs Regression
  / Nonliniarity in last layer}

Given the nature of the neural net the output of the last layer are
real numbers. For regression tasks this is desirable, for
classification problems however some transformations might be
necessary.
As the goal in the latter is to predict a certain class or classes for
an object the output needs to be of a form that allows this
interpretation.
Commonly the nodes in the output layer each correspond to a class and
the class chosen as prediction is the one with the highest value at
the corresponding output node.
The naive transformation to achieve this is transforming the output
vector $o$ into a one-hot vector
\[
  \text{pred}_i =
  \begin{cases}
    1,& \text{if } o_i = \max_j o_j \\
    0,& \text{else}.
  \end{cases}
\]
This however makes training the model with gradient based methods impossible, as the derivative of
the transformation is either zero or undefined.
A continuous transformation that is close to the argmax one is given by
softmax
\[
  \text{softmax}(o)_i = \frac{e^{o_i}}{\sum_j e^{o_j}}.
\]
The softmax function transforms the realm of the output to the interval $[0,1]$
and the individual values sum to one, thus the output can be interpreted as
a probability for each class given the input.
Additionally to being differentiable this allows for evaluataing the
cetainiy of a prediction, rather than just whether it is accurate.

\todo{vielleicht additiv invarianz}
% Another property that makes softmax attractive is the invariance to addition
% \[
%   \text{sofmax}(o) = \text{softmax}(o + c
% \]


% In order to properly interpret the output of a neural network and
% training it, depending on the problem it might be advantageous to
% transform the output form the last layer. Given the nature of the
% neural network the value at each output node is a real number. This is
% desirable for applications where the desired output is a real numbered
% vector (e.g. steering inputs for a autonomous car), however for
% classification problems it is desirable to transform this
% output. Often classification problems are modeled in such a way that
% each output node corresponds to a class. Then the output vector needs
% to be normalized in order to give a prediction. The naive approach is
% to transform the output vector $o$ into a one-hot vector $p$
% corresponding to a $0$
% entry for all classes except one, which is the predicted class.

% \[
%   p_i =
%   \begin{cases}
%     1,&  i < j, \forall i,j \in \text{arg}\max o_i, \\ 
%     0,& \text{else.}
%   \end{cases}
% \]\todo{besser formulieren}

% However this imposes difficulties in training the network as with this
% addition the model is no longer differentiable which imitates the
% ways the model can be trained. Additionally information about the
% ``certainty'' for each class in the prediction gets lost. A popular
% way to circumvent this problem is to normalize the output vector is
% such a way that the entries add up to one, this allows for the
% interpretation of probabilities assigned to each class.

\subsubsection{Error Measurement}

In order to make assessment about the quality of a network $\mathcal{NN}$ and train
it we need to discuss how we measure error. The choice of the error
function is highly dependent on the type of the problem. For
regression problems a commonly used error measure is the mean squared
error (MSE)
which for a function $f$ and data $(x_i,y_i), i=1,\dots,n$ is given by
\[
  MSE(f) = \frac{1}{n} \sum_i^n \left(f(x_i) - y_i\right)^2.
\]
However depending on the problem error measures with differnt
properties might be needed, for example in some contexts it is
required to consider a proportional rather than absolute error as is
common in time series models. \todo{komisch}

As discussed above the output of a neural network for a classification
problem can be interpreted as a probability distribution over the classes
conditioned on the input. In this case it is \todo{can?} desirable to
use error functions designed to compare probability distributions. A
widespread error function for this use case is the cross entropy (\textcite{PRML}),
which for two discrete distributions $p, q$ with the same realm $C$ is given by
\[
  H(p, q) = \sum_{c \in C} p(c) \ln\left(\frac{1}{q(c)}\right),
\]
which compares a $q$ to a true underlying distribution $p$.  
For a data set $(x_i,y_i), i = 1,\dots,n$ where each $y_{i,c}$
corresponds to the probability of class $c$ given $x_i$ and predictor
$f$ we get the loss function
\[
  Bla = \sum_{i=1}^n H(y_i, f(x_i)).
\]

-Maximum Likelihood
-Ableitung mit softmax pseudo linear -> fast improvemtns possible

\subsubsection{Gradient Descent Algorithm}

When trying to fit a neural network it is hard
to predict the impact of the single parameters on the accuracy of the
output. Thus applying numeric optimization algorithms is the only
feasible way to fit the model. A attractive algorithm for training
neural networks is gradient descent where each parameter $\theta_i$ is
iterative changed according to the gradient regarding the error
measure and a step size $\gamma$. For this all parameters are
initialized (often random or close to zero) and then iteratively
updated until a certain criteria is hit, mostly either being a fixed
number of iterations or a desired upper limit for the error measure.
% For a function $f_\theta$ with parameters $\theta \in \mathbb{R}^n$
% and a error function $L(f_\theta)$ the gradient descent algorithm is
% given in \ref{alg:gd}.

\begin{algorithm}[H]
  \SetAlgoLined
  \KwInput{function $f_\theta$ with parameters $\theta \in
    \mathbb{R}^n$ \newline step size $\gamma$}
  initialize $\theta^0$\;
  $i \leftarrow 1$\;
  \While{While termination condition is not met}{
    $\nabla \leftarrow \frac{\mathrm{d}f_\theta}{\mathrm{d} \theta}\vert_{\theta^{i-1}}$\;
    $\theta^i \leftarrow \theta^{i-1} - \gamma \nabla $\;
    $i \leftarrow i +1$\;
  }
  
  \caption{Gradient Descent}
  \label{alg:gd}
\end{algorithm}

The algorithm for gradient descent is given in
Algorithm~\ref{alg:gd}. In the context of fitting a neural network
$f_\theta$ corresponds to the error measurement of the network
$L\left(\mathcal{NN}_{\theta}\right)$ where $\theta$ is a vector
containing all the weights and biases of the network.
As ca be seen this requires computing the derivative of the network
with regard to each variable. With the number of variables getting
large in networks with multiple layers of high neuron count naively
computing these can get quite memory and computational expensive. But
by using the chain rule and exploiting the layered structure we can
compute the gradient much more efficiently by using backpropagation
first introduced by \textcite{backprop}.

\subsubsection{Backpropagation}

As with an increasing amount of layers the derivative of a loss
function with respect to a certain variable becomes more intensive to
compute there have been efforts in increasing the efficiency of
computing these derivatives. Today the BACKPROPAGATION algorithm is
widely used to compute the derivatives needed for the optimization
algorithms. Here instead of naively calculating the derivative for
each variable, the chain rule is used in order to compute derivatives
for each layer from output layer towards the first layer while only
needing to ....

\[
  \frac{\partial L(...)}{}
\]

%%% Local Variables:
%%% mode: latex
%%% TeX-master: "main"
%%% End: