\relax 
\@writefile{toc}{\contentsline {section}{\numberline {1}What Is an SVM? Margins and Support Vectors}{1}{}\protected@file@percent }
\newlabel{sec:what}{{1}{1}{}{section.1}{}}
\newlabel{sec:margins}{{1}{1}{}{section.1}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {1.1}The Margin Decides Where the Hyperplane Goes}{1}{}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {1.2}Functional and Geometric Margin}{2}{}\protected@file@percent }
\newlabel{sec:geomargin}{{1.2}{2}{}{subsection.1.2}{}}
\@writefile{toc}{\contentsline {section}{\numberline {2}The Optimal Margin Classifier and Its Dual}{3}{}\protected@file@percent }
\newlabel{sec:primal}{{2}{3}{}{section.2}{}}
\newlabel{sec:dual}{{2}{3}{}{section.2}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}The Primal Problem}{3}{}\protected@file@percent }
\newlabel{eq:primal}{{4}{3}{}{equation.4}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}The Lagrangian}{3}{}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {2.3}Minimize Over $\mathbf  {w},b$}{4}{}\protected@file@percent }
\newlabel{eq:wdual}{{6}{4}{}{equation.6}{}}
\newlabel{eq:balance}{{7}{4}{}{equation.7}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.4}The Dual Problem}{4}{}\protected@file@percent }
\newlabel{eq:dualhard}{{9}{4}{}{equation.9}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.5}Support Vectors}{4}{}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {3}Soft Margin: Non-Separable Data}{5}{}\protected@file@percent }
\newlabel{sec:soft}{{3}{5}{}{section.3}{}}
\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces One outlier and the hard margin. (a) Separable data: the hard-margin SVM, margin $1/\|\mathbf  {w}\|=0.84$, two support vectors (circled). (b) The same data plus one red point near the blue class: still separable, so the hard margin must accommodate it; the hyperplane tilts by about $44^\circ $ and the margin drops to $0.25$. Dotted gray: the boundary of (a). (c) The soft margin defined below ($C=0.3$) instead lets the outlier violate the margin (slack $\xi \approx 2.2$, misclassified) and recovers almost the boundary of (a), with a wide margin.}}{5}{}\protected@file@percent }
\newlabel{fig:outlier}{{1}{5}{}{figure.1}{}}
\newlabel{eq:softprimal}{{11}{5}{}{equation.11}{}}
\newlabel{eq:dualsoft}{{12}{6}{}{equation.12}{}}
\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces Linear soft-margin SVM on overlapping classes. Solid: decision boundary; dashed: margins $f(\mathbf  {x})=\pm 1$. Black circles: free support vectors ($0<\alpha _i<C$), exactly on the margin. Orange circles: support vectors at the bound ($\alpha _i=C$), inside the band or misclassified. All other points have $\alpha _i=0$. Small $C$ (left): wide margin, many violations, many support vectors. Large $C$ (right): narrow margin, few violations.}}{6}{}\protected@file@percent }
\newlabel{fig:linearC}{{2}{6}{}{figure.2}{}}
\@writefile{toc}{\contentsline {section}{\numberline {4}Kernels: The SVM in Feature Space}{7}{}\protected@file@percent }
\newlabel{sec:kernels}{{4}{7}{}{section.4}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {4.1}The Dual Needs Only $K$}{7}{}\protected@file@percent }
\newlabel{eq:dualkernel}{{13}{7}{}{equation.13}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {4.2}Predicting Without $\mathbf  {w}$}{7}{}\protected@file@percent }
\newlabel{eq:predict}{{14}{7}{}{equation.14}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {4.3}Choosing the Kernel and $C$}{7}{}\protected@file@percent }
\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces RBF-kernel SVM on the two-moons dataset. Rows: $\sigma =1$ (top) and $\sigma =0.25$ (bottom). Columns: $C=0.1,1,100$. Solid: decision boundary; dashed: margins; circled: support vectors. Small $\sigma $ and large $C$ (bottom right) fit the training data perfectly with a very irregular boundary.}}{8}{}\protected@file@percent }
\newlabel{fig:rbf}{{3}{8}{}{figure.3}{}}
\@writefile{toc}{\contentsline {section}{\numberline {5}Solving the Dual: SMO}{8}{}\protected@file@percent }
\newlabel{sec:smo}{{5}{8}{}{section.5}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Why a Special Algorithm}{8}{}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.2}One SMO Step}{9}{}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.3}Which Pairs? The Simplified SMO Algorithm}{10}{}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {5.4}Worked Example}{10}{}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {6}Multiclass SVMs}{10}{}\protected@file@percent }
\newlabel{sec:multiclass}{{6}{10}{}{section.6}{}}
\@writefile{toc}{\contentsline {section}{\numberline {7}Implementation Notes (HW3)}{11}{}\protected@file@percent }
\newlabel{sec:implementation}{{7}{11}{}{section.7}{}}
\gdef \@abspage@last{11}