add lots of stuff to materialandmethods and results

src/experimentalresults.tex

@@ -1,5 +1,6 @@
-\section{Experimental Results}
+\section{Experimental Results}\label{sec:experimental-results}
-\subsection{Does Active-Learning benefit the learning process?}
+
+\subsection{Does Active-Learning benefit the learning process?}\label{subsec:does-active-learning-benefit-the-learning-process?}
 
 With the test setup described in~\ref{sec:implementation} a test series was performed.
 Several different batch sizes $\mathcal{B} = \left\{ 2,4,6,8 \right\}$ and sample sizes $\mathcal{S} = \left\{ 2\mathcal{B}_i,4\mathcal{B}_i,5\mathcal{B}_i,10\mathcal{B}_i \right\}$
@@ -8,57 +9,57 @@ We define the baseline (passive learning) AUC curve as the supervised learning p
 The following graphs are only a subselection of the test series which give the most insights.
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_2_10}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_2_10}
     \caption{AUC with $\mathcal{B} = 2$ and $\mathcal{S}=10$}
+    \label{fig:auc_normal_lowcer_2_10}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_2_20}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_2_20}
     \caption{AUC with $\mathcal{B} = 2$ and $\mathcal{S}=20$}
+    \label{fig:auc_normal_lowcer_2_20}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_2_50}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_2_50}
     \caption{AUC with $\mathcal{B} = 2$ and $\mathcal{S}=50$}
+    \label{fig:auc_normal_lowcer_2_50}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_4_16}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_4_16}
     \caption{AUC with $\mathcal{B} = 4$ and $\mathcal{S}=16$}
+    \label{fig:auc_normal_lowcer_4_16}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_4_24}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_4_24}
     \caption{AUC with $\mathcal{B} = 4$ and $\mathcal{S}=24$}
+    \label{fig:auc_normal_lowcer_4_24}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_8_16}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_8_16}
     \caption{AUC with $\mathcal{B} = 8$ and $\mathcal{S}=16$}
+    \label{fig:auc_normal_lowcer_8_16}
 \end{figure}
 
 \begin{figure}
-    \label{fig:auc_normal_lowcer_8_32}
     \centering
     \hspace*{-0.1\linewidth}\includegraphics[width=1.2\linewidth]{../rsc/AUC_normal_lowcer_8_32}
     \caption{AUC with $\mathcal{B} = 8$ and $\mathcal{S}=32$}
+    \label{fig:auc_normal_lowcer_8_32}
 \end{figure}
 
-Generally a pattern can be seen: The lower the batch size the more benefits are gained by active learning.
+Generally a pattern can be seen: The lower the batch size $\mathcal{B}$ the more benefits are gained by active learning.
 This may be caused by the fast model convergence.
-The lower the batch size the more pre-prediction decision points are required.
+The lower $\mathcal{B}$ the more pre-prediction decision points are required.
 This helps directing the learning with better samples of the selected metric.
 When the batch size is higher the model already converges to a good AUC value before the same amount of pre-predictions is reached.
 
@@ -66,17 +67,31 @@ Moreover, when increasing the sample-space $\mathcal{S}$ from where the pre-pred
 This is because the selected subset $\pmb{x} \sim \mathcal{X}_U$ has a higher chance of containing relevant elements corresponding to the selected metric.
 But keep in mind this improvement comes with a performance penalty because more model evaluations are required to predict the ranking scores.
 
-% todo
+\ref{fig:auc_normal_lowcer_2_10};\ref{fig:auc_normal_lowcer_2_20};\ref{fig:auc_normal_lowcer_2_50} shows the AUC curve with a batch size of 2 and a sample size of 10, 20, 50 respectively.
-\ref{fig:auc_normal_lowcer_2_10} shows the AUC curve with a batch size of 2 and a sample size of 10.
+On all three graphs the active learning curve outperforms the passive learning curve in all four scenarios.
-Todo add some references to the graphs.
+Generally the higher the sample space $\mathcal{S}$ the better the performance.
 
+\ref{fig:auc_normal_lowcer_4_16};\ref{fig:auc_normal_lowcer_4_24} shows the AUC curve with a batch size of 4 and a sample size of 16, 24 respectively.
+The performance is already much worse compared to the results from above with a batch size of 2.
+Only the low certainty first approach outperforms the passive learning in both cases.
+The other methods are as good or worse than the passive learning curve.
+
+\ref{fig:auc_normal_lowcer_8_16};\ref{fig:auc_normal_lowcer_8_32} shows the AUC curve with a batch size of 8 and a sample size of 16, 32 respectively.
+The performance is even worse compared to the results from above with a batch size of 4.
+This might be the case because the model already converges to a good AUC value before the same amount of pre-predictions is reached.
 
 \subsection{Is Dagster and Label-Studio a proper tooling to build an AL
 Loop?}\label{subsec:is-dagster-and-label-studio-a-proper-tooling-to-build-an-al
 loop?}
 
 The combination of Dagster and Label-Studio is a good choice for building an active-learning loop.
+Dagster provides a clean way to build pipelines and to keep track of the data in the Web UI\@.
+Label-Studio provides a great api which can be used to update the predictions of the model from the dagster pipeline.
+
+% todo write stuff here
 
 \subsection{Does balancing the learning samples improve performance?}\label{subsec:does-balancing-the-learning-samples-improve-performance?}
 
 Not really.
+
+% todo add img and add stuff

src/implementation.tex

@@ -1,29 +1,5 @@
 \section{Implementation}\label{sec:implementation}
 
-\subsection{Dagster with Label-Studio}\label{subsec:dagster-with-label-studio}
-
-The main goal is to implement an active learning loop with the help of Dagster and Label-Studio.
-The active learning loop was split as much as possible into assets and graph assets.
-This helps building reusable building blocks and to keep the code clean.
-
-\begin{figure}
-    \centering
-    \includegraphics[width=\linewidth]{../rsc/dagster/assets}
-    \caption{Dagster asset graph}
-\end{figure}
-
-\begin{figure}
-    \centering
-    \subfloat[Score prediction graph asset]{
-        \includegraphics[width=0.45\linewidth]{../rsc/dagster/predict_scores}
-    }
-    \hfill
-    \subfloat[Model training graph asset]{
-        \includegraphics[width=0.45\linewidth]{../rsc/dagster/train_model}
-    }
-    \caption{Dagster graph assets}
-\end{figure}
-
 \subsection{Jupyter}\label{subsec:jupyter}
 
 To get accurate performance measures the active-learning process was implemented in a Jupyter notebook first.
@@ -58,3 +34,48 @@ match predict_mode:
 \end{lstlisting}
 
 Moreover, the Dataset was manually imported and preprocessed with random augmentations.
+
+\subsection{Dagster with Label-Studio}\label{subsec:dagster-with-label-studio}
+
+The main goal is to implement an active learning loop with the help of Dagster and Label-Studio.
+The active learning loop was split as much as possible into assets and graph assets.
+This helps building reusable building blocks and to keep the code clean.
+
+Most of the python routines implemented in section~\ref{subsec:jupyter} were reused here and just slightly modified to fit the Dagster pipeline.
+
+% todo short this figure to half!
+\begin{figure}
+    \centering
+    \includegraphics[width=\linewidth]{../rsc/dagster/assets}
+    \caption{Dagster asset graph}
+    \label{fig:dagster_assets}
+\end{figure}
+
+\ref{fig:dagster_assets} shows the implemented assets in which the task is split.
+Whenever a asset materializes it is stored in the Dagster database.
+This helps to keep track of the data and to rerun the pipeline with the same data.
+\textit{train\_sup\_model} is the main asset that trains the model with the labeled samples.
+\textit{inference\_unlabeled\_samples} is the asset that predicts the scores for the unlabeled samples und updates them with the Label-Studio API.
+This pipeline is triggered by the machine learning backend defined within Label-Studio.
+It runs every $\mathcal{B} - t$ samples, where t is a buffer to compensate the time the pipeline needs to run to avoid overlaps (e.g.\ t=10).
+Such that Label-Studio has always samples with scores produced by the current model to rank the samples presented to the user.
+\begin{figure}
+    \centering
+    \subfloat[Score prediction graph asset]{
+        \includegraphics[width=0.45\linewidth]{../rsc/dagster/predict_scores}
+        \label{fig:predict_scores}
+    }
+    \hfill
+    \subfloat[Model training graph asset]{
+        \includegraphics[width=0.45\linewidth]{../rsc/dagster/train_model}
+    \label{fig:train_model}
+    }
+    \caption{Dagster graph assets}
+\end{figure}
+
+\ref{fig:train_model} shows the train model asset in detail.
+It loads the data for training, trains the model and saves the model automatically due to the Dagster asset system.
+Moreover, testing data is loaded and the model is evaluated with the test data to get a performance measure.
+\ref{fig:predict_scores} shows the predict scores asset in detail.
+It samples $\mathcal{S}$ samples from the unlabeled samples $\mathcal{X}_U$ and predicts the scores.
+Then it connects to the Label-Studio API with an API key and updates the scores of the samples.

src/introduction.tex

@@ -2,9 +2,10 @@
 \subsection{Motivation}\label{subsec:motivation}
 For most supervised learning tasks lots of training samples are essential.
 With too less training data the model will not generalize well and not fit a real world task.
-Labeling datasets is commonly seen as an expensive task and wants to be avoided as much as possible.
+Labeling datasets is commonly seen as an expensive task and wants to be avoided as much as possible.\cite{generalAI}
 That's why there is a machine-learning field called active learning.
-The general approach is to train a model that predicts within every iteration a ranking metric or Pseudo-Labels which then can be used to rank the importance of samples to be labeled.
+The general approach is to train a model that predicts within every iteration a ranking metric or Pseudo-Labels which then can be used to rank the importance of samples to be labeled by an oracle.
+These labeled are then used to train the model.\cite{activelearning}
 
 The goal of this practical work is to test active learning within a simple classification task and evaluate its performance.
 \subsection{Research Questions}\label{subsec:research-questions}
@@ -20,5 +21,11 @@ The sample-selection metric might select samples just from one class by chance.
 Does balancing this distribution help the model performance?
 \subsection{Outline}\label{subsec:outline}
 
+In section~\ref{sec:material-and-methods} we talk about general methods and materials we use.
+First the problem is modeled mathematically in~\ref{subsubsec:mathematicalmodeling} and then implemented and benchmarked in a Jupyter notebook~\ref{subsubsec:jupyternb}
+Section~\ref{sec:implementation} gives deeper insights to the implementation for the interested reader.
+The conclusion~\ref{subsec:conclusion} provides a overview of the findings, highlighting the benefits of active learning.
+Additionally the outlook section suggests avenues for future research which are not covered in this work.
+The experimental results are well-presented with clear figures illustrating the performance of active learning across different sample sizes and batch sizes.
 
-Todo talk about what we do in which section
+% todo proper linking to sections

src/materialandmethods.tex

@@ -18,9 +18,24 @@ An Op is a function that performs a task and can be used to split the code into
 Dagster has a well-built web interface to monitor jobs and pipelines. \cite{dagster}
 
 \subsubsection{Label-Studio}
-\subsubsection{Pytorch}
 
-\subsubsection{Imagenet}
+Label-Studio is a data labeling tool that can be used to label images, text, audio and video data.
+Which makes it an excellent choice incorporating human feedback into an active learning loop.
+
+Label Studio provides a wide range of annotation interfaces and can be extended with custom ones.
+Any arbitrary data can be passed to the labelling frontend using labelling tasks in the form of json files.
+It is open-source and can be used for free.
+Label Studio offers a seamless integration with active learning pipelines by allowing to define custom machine-learning backends.
+It is designed for scalability and can be easily deployed on a cloud infrastructure using Kubernetes or Helm.\cite{labelstudio}
+
+\subsubsection{Jupyter Notebook}\label{subsubsec:jupyternb}
+
+A Jupyter notebook is a shareable document which combines code and its output, text and visualizations.
+The notebook along with the editor provides a environment for fast prototyping and data analysis.
+It is widely used in the data science, mathematics and machine learning community.
+
+In the case of this practical work it can be used to test and evaluate the active learning loop before implementing it in a Dagster pipeline. \cite{jupyter}
+
 \subsubsection{Muffin vs chihuahua}
 Muffin vs chihuahua is a free dataset available on Kaggle.
 It consists of $\sim6000$ images of muffins and chihuahuas.
@@ -34,15 +49,13 @@ It is used in this practical work for a binary classification task to evaluate t
 Active learning is a subfield of supervised learning.
 The key idea is if the algorithm is allowed to choose the data it learns from, it can perform better with less data.
 A supervised classifier requires hundreds or even thousands of labeled samples to perform well.
-Those labeled samples must be manually labeled by an oracle (human expert).
+Those labeled samples must be manually labeled by an oracle (human expert).\cite{RubensRecSysHB2010}
 
 Clearly this results in a huge bottleneck for the training procedure.
-Active learning aims to overcome this bottleneck by selecting the most informative samples to be labeled.
+Active learning aims to overcome this bottleneck by selecting the most informative samples to be labeled.\cite{settles.tr09}
-
-Todo \cite{RubensRecSysHB2010} \cite{settles.tr09}
 
+The active learning process can be modeled as a loop as shown in~\ref{fig:active-learning-workflow}.
 \begin{figure}
-    \label{fig:active-learning-workflow}
     \centering
     \begin{tikzpicture}[node distance=2cm]
         \node (start) [startstop] {Start};
@@ -64,8 +77,14 @@ Todo \cite{RubensRecSysHB2010} \cite{settles.tr09}
         \draw [arrow] (pro4) -- (pro2);
     \end{tikzpicture}
     \caption{Basic active-learning workflow}
+    \label{fig:active-learning-workflow}
 \end{figure}
 
+The active learning loop starts with the model inference on $\mathcal{S}$ samples.
+The most uncertain samples of size $\mathcal{B}$ are selected and given to the oracle\footnote{Human annotator} for labeling.
+Those labeled samples are then used to train the model.
+The loop starts again with the new model and draws new samples from the unlabeled sample set $\mathcal{X}_U$.
+
 \subsubsection{Semi-Supervised learning}
 In traditional supervised learning we have a labeled dataset.
 Each datapoint is associated with a corresponding target label.
@@ -90,13 +109,24 @@ The more the curve ascents the upper-left or bottom-right corner the better the
 \begin{figure}
     \centering
     \includegraphics[width=\linewidth/2]{../rsc/Roc_curve.svg}
-    \caption{Architecture convolutional neural network. Image by \href{https://cointelegraph.com/explained/what-are-convolutional-neural-networks}{SKY ENGINE AI}}
+    \caption{ROC curve comparision of two classifiers. \cite{ROCWikipedia}}
     \label{fig:roc-example}
 \end{figure}
 
-Furthermore, the area under this curve is called AUR curve and a useful metric to measure the performance of a binary classifier.
+Furthermore, the area under this curve is called AUR curve and a useful metric to measure the performance of a binary classifier. \cite{suptechniques}
 
 \subsubsection{RESNet}
+
+Residual neural networks are a special type of neural network architecture.
+They are especially good for deep learning and have been used in many state-of-the-art computer vision tasks.
+The main idea behind ResNet is the skip connection.
+The skip connection is a direct connection from one layer to another layer which is not the next layer.
+This helps to avoid the vanishing gradient problem and helps with the training of very deep networks.
+ResNet has proven to be very successful in many computer vision tasks and is used in this practical work for the classification task.
+There are several different ResNet architectures, the most common are ResNet-18, ResNet-34, ResNet-50, ResNet-101 and ResNet-152. \cite{resnet}
+
+Since the dataset is relatively small and the two class classification task is relatively easy the ResNet-18 architecture is used in this practical work.
+
 \subsubsection{CNN}
 Convolutional neural networks are especially good model architectures for processing images, speech and audio signals.
 A CNN typically consists of Convolutional layers, pooling layers and fully connected layers.
@@ -139,9 +169,7 @@ And~\eqref{eq:crelbinary} is the special case of the general Cross Entropy Loss
 
 $\mathcal{L}(p,q)$~\eqref{eq:crelbinarybatch} is the Binary Cross Entropy Loss for a batch of size $\mathcal{B}$ and used for model training in this PW.
 
-\subsubsection{Adam}
+\subsubsection{Mathematical modeling of problem}\label{subsubsec:mathematicalmodeling}
-
-\subsubsection{Mathematical modeling of problem}
 
 Here the task is modeled as a mathematical problem to get a better understanding of how the problem is solved.
 
@@ -188,10 +216,10 @@ We define $\text{min}_n(S)$ and $\text{max}_n(S)$ respectively in~\ref{eq:minnot
 
 This notation helps to define which subsets of samples to give the user for labeling.
 There are different ways how this subset can be chosen.
-In this PW we do the obvious experiments with High-Certainty first~\ref{subsec:low-certainty-first}, Low-Certainty first~\ref{subsec:high-certainty-first}.
+In this PW we do the obvious experiments with High-Certainty first~\ref{par:low-certainty-first}, Low-Certainty first~\ref{par:high-certainty-first}.
 Furthermore, the two mixtures between them, half-high and half-low certain and only the middle section of the sorted certainty scores.
 
-\paragraph{Low certainty first}
+\paragraph{Low certainty first}\label{par:low-certainty-first}
 We take the samples with the lowest certainty score first and give it to the user for labeling.
 This is the most intuitive way to do active learning and might also be the most beneficial.
 
@@ -199,7 +227,7 @@ This is the most intuitive way to do active learning and might also be the most
     \mathcal{X}_t  = \text{min}_\mathcal{B}(S(z))
 \end{equation}
 
-\paragraph{High certainty first}
+\paragraph{High certainty first}\label{par:high-certainty-first}
 We take the samples with the highest certainty score first and give it to the user for labeling.
 The idea behind this is that the model is already very certain about the prediction and the user can confirm this.
 This might help ignoring labels which are irrelevant for the model.

src/sources.bib

@@ -58,6 +58,22 @@ and Sardinha, Alberto",
     note = "[Online; accessed 12-April-2024]"
 }
 
+@misc{labelstudio,
+    author = {},
+    title = {{Label Studio Documentation}},
+    howpublished = "\url{https://labelstud.io/guide/}",
+    year = {2024},
+    note = "[Online; accessed 13-May-2024]"
+}
+
+@misc{jupyter,
+    author = {},
+    title = {{Project Jupyter Documentation}},
+    howpublished = "\url{https://docs.jupyter.org/en/latest/}",
+    year = {2024},
+    note = "[Online; accessed 13-May-2024]"
+}
+
 @misc{muffinsvschiuahuakaggle,
     author = {},
     title = {{Muffin vs Chihuahua Kaggle Dataset}},
@@ -85,3 +101,37 @@ doi = {10.1007/978-0-387-85820-3_23}
     Type = {Computer Sciences Technical Report},
     Year = {2009},
 }
+
+@misc{generalAI,
+    author        = {Johannes Brandstetter},
+    title         = {Lecture notes in Theoretical Concepts of Machine Learning},
+    month         = {May},
+    year          = {2024},
+    publisher={Johannes Kepler Universität Linz}
+}
+
+@misc{suptechniques,
+    author        = {Andreas Radler, Markus Holzleitner},
+    title         = {Lecture notes in Machine Learning: Supervised Techniques},
+    month         = {September},
+    year          = {2022},
+    publisher={Johannes Kepler Universität Linz}
+}
+
+@online{ROCWikipedia,
+    author  = "Wikimedia Commons",
+    title   = "Receiver operating characteristic",
+    year    = "2024",
+    urlseen = "13-05-24",
+    url     = "https://commons.wikimedia.org/wiki/File:Roc_curve.svg",
+    note    = "File: \ttfamily{Roc curve.svg}",
+}
+
+@misc{resnet,
+    title={Deep Residual Learning for Image Recognition},
+    author={Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
+    year={2015},
+    eprint={1512.03385},
+    archivePrefix={arXiv},
+    primaryClass={cs.CV}
+}