Martin Thoma преди 11 години
родител
ревизия
1c329237ae

BIN
presentations/Bachelor-Final-Presentation/LaTeX/Bachelor-Final-Presentation.pdf


+ 4 - 10
presentations/Bachelor-Final-Presentation/LaTeX/Bachelor-Final-Presentation.tex

@@ -26,22 +26,16 @@
 \section{What is my Bachelor's thesis about?}
 \input{introduction}
 
-\section{write-math.com and HWRT}
-\input{write-math}
-
-% \section{Preprocessing and Features}
-% \input{preprocessing}
-% \input{features}
+\section{Preprocessing and Features}
+\input{preprocessing}
+\input{features}
 
 \section{Evaluation}
 \input{evaluation}
 
-% \section{What will I do next?}
-% \input{will-do}
-
 \section*{End}
 \subsection{End}
-\input{sources}
+\input{end}
 \framedgraphic{Thanks for Your Attention!}{../images/xi.png}
 
 \end{document}

+ 25 - 0
presentations/Bachelor-Final-Presentation/LaTeX/end.tex

@@ -0,0 +1,25 @@
+\subsection{HWRT and write-math.com}
+\begin{frame}{HWRT and write-math.com}
+    Two software projects were created:
+    \begin{itemize}
+    \item \href{http://write-math.com}{write-math.com}: A website where
+          on-line handwritten data gets collected and classified
+    \item \href{https://github.com/MartinThoma/hwrt}{hwrt}: The
+          \textit{handwriting recognition toolkit} is a Python project for
+          handwriting recognition
+    \end{itemize}
+
+    This presentation and the bachelor's thesis will be at
+    \href{http://martin-thoma.com/write-math/}{martin-thoma.com/write-math}.
+\end{frame}
+
+\subsection{Sources}
+\begin{frame}{Image Sources}
+    \begin{itemize}
+	\item \href{https://commons.wikimedia.org/wiki/File:Server-multiple.svg}{Server} by RRZEicons
+    \item \href{https://commons.wikimedia.org/wiki/File:Computer-aj_aj_ashton_01.svg}{Desktop Computer} by Ed g2s,
+          Ironbrother, Kierancassel and Msgj
+    \item \href{https://commons.wikimedia.org/wiki/File:Server_by_mimooh.svg}{Server} by Mimooh
+    \end{itemize}
+\end{frame}
+

+ 6 - 5
presentations/Bachelor-Final-Presentation/LaTeX/evaluation.tex

@@ -92,17 +92,18 @@
 \begin{frame}{Complex classifier}
     \textbf{Preprocessing:} Connect strokes, scale, shift and linear interpolation\\
     \textbf{Features:} Coordinates of 80 points (4 strokes with 20 points each), re-curvature per stroke, ink, stroke count, aspect ratio\\
-    \textbf{Learning:} MLP, 1000 epochs, LR $\eta=0.1$, Momentum $\alpha=0.1$
+    \textbf{Learning:} MLP, 1000 epochs, LR $\eta=0.1$, Momentum $\alpha=0.1$, supervised layer-wise pretraining
 \begin{table}[htb]
     \centering
     \begin{tabular}{lrrrrrr}
     \toprule
-    \multirow{2}{*}{System}  & \multicolumn{3}{c}{Classification error}\\ 
+    \multirow{2}{*}{System}  & \multicolumn{6}{c}{Classification error}\\
+    \cmidrule(l){2-7}
               & TOP1                   & change                 & TOP3                  & change                 & MER                   & change \\\midrule
     $B_{1,c}$ & $\SI{20.96}{\percent}$ & $\SI{-2.38}{\percent}$ & $\SI{5.24}{\percent}$ & $\SI{-1.56}{\percent}$ & $\SI{5.13}{\percent}$ & $\SI{-1.51}{\percent}$ \\
-    $B_{2,c}$ & $\SI{20.10}{\percent}$ & $\SI{-1.41}{\percent}$ & $\SI{4.44}{\percent}$ & $\SI{-1.31}{\percent}$ & $\SI{4.36}{\percent}$ & $\SI{-1.31}{\percent}$ \\
-    $B_{3,c}$ & $\SI{21.51}{\percent}$ & $\SI{-0.42}{\percent}$ & $\SI{4.89}{\percent}$ & $\SI{-0.85}{\percent}$ & $\SI{4.80}{\percent}$ & $\SI{-0.84}{\percent}$ \\
-    $B_{4,c}$ & $\SI{00.00}{\percent}$ & $\SI{-0.00}{\percent}$ & $\SI{0.00}{\percent}$ & $\SI{-0.00}{\percent}$ & $\SI{0.00}{\percent}$ & $\SI{-0.00}{\percent}$ \\
+    $B_{2,c}$ & $\SI{18.26}{\percent}$ & $\SI{-3.25}{\percent}$ & $\SI{4.07}{\percent}$ & $\SI{-1.68}{\percent}$ & \underline{$\SI{3.98}{\percent}$} & $\SI{-1.69}{\percent}$ \\
+    $B_{3,c}$ & \underline{$\SI{18.19}{\percent}$} & $\SI{-3.74}{\percent}$ & \underline{$\SI{4.06}{\percent}$} & $\SI{-1.68}{\percent}$ & $\SI{3.99}{\percent}$ & $\SI{-1.65}{\percent}$ \\
+    $B_{4,c}$ & $\SI{18.57}{\percent}$ & $\SI{-5.31}{\percent}$ & $\SI{4.25}{\percent}$ & $\SI{-1.87}{\percent}$ & $\SI{4.18}{\percent}$ & $\SI{-1.86}{\percent}$ \\
     \bottomrule
     \end{tabular}
     \caption{Error rates of the complex recognizer systems.}

+ 2 - 0
presentations/Bachelor-Final-Presentation/LaTeX/features.tex

@@ -18,6 +18,8 @@
             \item Center point
             \item Bitmap
             \item Bounding box (width, height, time)
+            \item Re-curvature
+            \item Ink
         \end{itemize}
     \end{itemize}
 \end{frame}

+ 2 - 12
presentations/Bachelor-Final-Presentation/LaTeX/introduction.tex

@@ -5,7 +5,7 @@
         \item Recognition of handwritten mathematical symbols
         \item On-line recognition, not OCR!
         \item Given a series of points $(x(t), y(t), b(t))$\\
-              I want to get the proper \LaTeX{} command.
+              I want to get the \LaTeX{} command.
     \end{itemize}
 \end{frame}
 
@@ -16,15 +16,5 @@
         \item It's much harder to find complete formulas.
     \end{itemize}
 
-    % I want to
-    % \begin{itemize}
-    %     \item provide a tool that enables beginners to get the best \LaTeX{} command
-    %           for their formula,
-    %     \item find out what works best for symbol recognition
-    %     \item and provide data and a platform to test new ideas for classifiers
-    % \end{itemize}
-
-    For now: recognition of isolated symbols. That means:
-
-    single symbol \enquote{formulas} rather than multi-symbol formulas
+    For now: recognition of isolated symbols.
 \end{frame}

+ 0 - 12
presentations/Bachelor-Final-Presentation/LaTeX/sources.tex

@@ -1,12 +0,0 @@
-\subsection{Sources}
-\begin{frame}{Image Sources}
-    \begin{itemize}
-	\item \href{https://commons.wikimedia.org/wiki/File:Server-multiple.svg}{Server} by RRZEicons
-    \item \href{https://commons.wikimedia.org/wiki/File:Computer-aj_aj_ashton_01.svg}{Desktop Computer} by Ed g2s,
-          Ironbrother, Kierancassel and Msgj
-    \item \href{https://commons.wikimedia.org/wiki/File:Server_by_mimooh.svg}{Server} by Mimooh
-    \end{itemize}
-
-    The presentation can be found at \url{http://tinyurl.com/write-math-short-presentation}
-\end{frame}
-

+ 0 - 24
presentations/Bachelor-Final-Presentation/LaTeX/will-do.tex

@@ -1,24 +0,0 @@
-\subsection{What will I do next?}
-\begin{frame}{What will I do next?}
-    \begin{itemize}
-        \item Include the currently best model in write-math.com
-        \item Evaluate preprocessing steps
-        \item Try other features
-        \item Try other topologies / trainings (e.g. pretraining, newbob)
-        \item Eventually try convolutional neural nets
-    \end{itemize}
-\end{frame}
-
-% \subsection{Far future}
-% \begin{frame}{What could be done?}
-%     \begin{itemize}
-%         \item Make use of audio data in a multimodal approach\\
-%               e.g. $R$ and $\mathcal{R}$
-%         \item Currently, the Lecture Translation system doesn't recognize math.\\
-%               You get \enquote{integral of e raised to the power of x d x} instead
-%               of $\int e^x \mathrm{d} x$.
-%         \item Spoken math is ambigous: $\sqrt{a+b}$ vs. $\sqrt{a} + b$
-%         \item The language model I create could help to find probable formulas
-%         \item The platform could be used to get more input data of users
-%     \end{itemize}
-% \end{frame}