lukas/Seminar_in_AI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

finish presentation

Browse Source
This commit is contained in:
lukas-heiligenbrunner 2023-03-24 14:17:21 +01:00
parent 2267d43209
commit 63bb67d77a
1 changed files with 20 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
presentation/main.tex
View File

@ -64,6 +64,7 @@
%---------------------------------------------------------------------------------------- %----------------------------------------------------------------------------------------
% todo pic of action % todo pic of action
\section{The Goal} \section{The Goal}
\begin{frame}{The goal} \begin{frame}{The goal}
\begin{itemize} \begin{itemize}
@ -105,7 +106,7 @@
\begin{itemize} \begin{itemize}
\item YouTube \item YouTube
\end{itemize} \end{itemize}
\item using semi-supervised learning might be benefitial \item Using semi-supervised learning might be benefitial
\end{itemize} \end{itemize}
\end{frame} \end{frame}
@ -143,11 +144,10 @@
\item Pseudo-labeling \item Pseudo-labeling
\item Train model on labeled data \item Train model on labeled data
\begin{itemize} \begin{itemize}
\item Eg. 1\% of data labeled \item Eg. 1\%/10\% of data labeled
\end{itemize} \end{itemize}
\item Confidence of prediction \item Confidence of prediction [Threshold]
\item If high enough \item Use pseudo-labels to predict unlabeled data
\item Use to predict unlabeled data
\end{itemize} \end{itemize}
\framebreak \framebreak
@ -174,11 +174,13 @@
\begin{itemize} \begin{itemize}
\item lower capacity \item lower capacity
\item better captures temporal dynamics in recognizing actions \item better captures temporal dynamics in recognizing actions
\item scene changes/motion over time
\end{itemize} \end{itemize}
\item Large model \item Large model
\begin{itemize} \begin{itemize}
\item better learns spatial semantics \item better learns spatial semantics
\item to distinguish different action instances \item to distinguish different action instances
\item localize/identify objects in specific scene
\end{itemize} \end{itemize}
\end{itemize} \end{itemize}
@ -190,7 +192,7 @@
\item Supplemented by lightweight auxiliary network \item Supplemented by lightweight auxiliary network
\begin{itemize} \begin{itemize}
\item Different structure \item Different structure
\item Fewer channels \item Fewer channels (smaller)
\end{itemize} \end{itemize}
\item Different representation of data complements primary backbone \item Different representation of data complements primary backbone
\end{itemize} \end{itemize}
@ -201,6 +203,10 @@
\end{frame} \end{frame}
\begin{frame}{Performance Perspectives} \begin{frame}{Performance Perspectives}
\begin{itemize}
\item 1\% labeled data + 400 Labels
\item Kinetics-400 dataset
\end{itemize}
\includegraphics[scale=.205]{rsc/performance_comparison} \includegraphics[scale=.205]{rsc/performance_comparison}
\end{frame} \end{frame}
@ -219,24 +225,24 @@
\begin{frame}[allowframebreaks]{How existing method \textit{FixMatch} works} \begin{frame}[allowframebreaks]{How existing method \textit{FixMatch} works}
\begin{itemize} \begin{itemize}
\item $\mathbbm{1} \coloneqq \text{Indicator Function}$
\item $B_u \coloneqq \text{Batchsize}$ \item $B_u \coloneqq \text{Batchsize}$
\item $\mathcal{T} \coloneqq \text{Confidence Threshold}$ \item $\tau \coloneqq \text{Confidence Threshold (Hyperparameter)}$
\item $F(\mathcal{T}_{\text{strong}}(u_i)) \coloneqq \text{Class distribution}$ \item $F(\mathcal{T}_{\text{strong}}(u_i)) \coloneqq \text{Class distribution}$
\item $p_i \coloneqq F(\mathcal{T}_{\text{weak}}(u_i))$ \item $p_i \coloneqq F(\mathcal{T}_{\text{weak}}(u_i))$
\item $\hat{y}_i \coloneqq \arg \max(p_i) \coloneqq \text{Pseudo Label}$ \item $\hat{y}_i \coloneqq \arg \max(p_i) \coloneqq \text{Pseudo Label}$
\item $\mathcal{H} \coloneqq \text{Cross-entropy loss}$ \item $\mathcal{H} \coloneqq \text{Cross-entropy loss}$
\item $\mathcal{L}_u \coloneqq \text{Loss on the unlabeled data}$ \item $\mathcal{L}_u \coloneqq \text{Loss on the unlabeled data}$
\item $\mathcal{F} \coloneqq \text{Model}$ \item $F \coloneqq \text{Model}$
\item $\mathbbm{1} \coloneqq \text{Indicator Function}$
\end{itemize} \end{itemize}
\begin{align*} \begin{align*}
\mathcal{L}_u = \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i) \geq \mathcal{T}) \mathcal{H}(\hat{y}_i,F(\mathcal{T}_{\text{strong}}(u_i))) \mathcal{L}_u = \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i) \geq \tau) \mathcal{H}(\hat{y}_i,F(\mathcal{T}_{\text{strong}}(u_i)))
\end{align*} \end{align*}
\framebreak \framebreak
\begin{itemize} \begin{itemize}
\item $\mathbbm{1}(\max(p_i) \geq \mathcal{T})$ \item $\mathbbm{1}(\max(p_i) \geq \tau)$
\begin{itemize} \begin{itemize}
\item 'confidence-based masking' \item 'confidence-based masking'
\item retain label only if largest probability is above threshold \item retain label only if largest probability is above threshold
@ -268,8 +274,8 @@
\begin{itemize} \begin{itemize}
\item Learning on unlabeled data \item Learning on unlabeled data
\begin{align*} \begin{align*}
\mathcal{L}_u^F &= \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i^A) \geq \mathcal{T}) \mathcal{H}(\hat{y}_i^A,F(\mathcal{T}_{\text{strong}}(u_i)))\\ \mathcal{L}_u^F &= \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i^A) \geq \tau) \mathcal{H}(\hat{y}_i^A,F(\mathcal{T}_{\text{strong}}(u_i)))\\
\mathcal{L}_u^A &= \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i^F) \geq \mathcal{T}) \mathcal{H}(\hat{y}_i^F,A(\mathcal{T}_{\text{strong}}(u_i)))\\ \mathcal{L}_u^A &= \frac{1}{B_u} \sum_{i=1}^{B_u} \mathbbm{1}(\max(p_i^F) \geq \tau) \mathcal{H}(\hat{y}_i^F,A(\mathcal{T}_{\text{strong}}(u_i)))\\
\end{align*} \end{align*}
\item Complete training objective \item Complete training objective
\begin{align*} \begin{align*}
@ -279,6 +285,7 @@
\end{itemize} \end{itemize}
\end{frame} \end{frame}
\section{Implementation} \section{Implementation}
\begin{frame}{Networks} \begin{frame}{Networks}