Indexed text attachment
differential-topology-of-parametrized-neural-networks.tex
Source for Differential Topology of Parametrized Neural Networks
View paper · Download original file
Lines 41–140 of 1074. Source text is displayed literally.
41 {\color{paperaccent}\thesection}{.55em}{}42\titlespacing*{\section}{0pt}{14pt plus 3pt minus 2pt}{5pt}43\titleformat{\subsection}{\normalsize\bfseries\raggedright}44 {\color{paperaccent}\thesubsection}{.55em}{}45\titlespacing*{\subsection}{0pt}{10pt plus 2pt minus 1pt}{4pt}46\captionsetup{font=small,labelfont={bf,color=paperaccent}}47 48\pagestyle{fancy}49\fancyhf{}50\fancyhead[L]{\footnotesize\sffamily\color{papermuted}Differential topology of parametrized neural networks}51\fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}52\renewcommand{\headrulewidth}{0pt}53\fancypagestyle{plain}{%54 \fancyhf{}%55 \fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}%56 \renewcommand{\headrulewidth}{0pt}%57}58 59\theoremstyle{plain}60\newtheorem{theorem}{Theorem}[section]61\newtheorem{proposition}[theorem]{Proposition}62\newtheorem{lemma}[theorem]{Lemma}63\newtheorem{corollary}[theorem]{Corollary}64\theoremstyle{definition}65\newtheorem{definition}[theorem]{Definition}66\newtheorem{example}[theorem]{Example}67\theoremstyle{remark}68\newtheorem{remark}[theorem]{Remark}69\DeclareMathOperator{\Map}{Map}70\DeclareMathOperator{\Tot}{Tot}71\DeclareMathOperator{\Fun}{Fun}72\DeclareMathOperator{\Hom}{Hom}73\DeclareMathOperator{\grad}{grad}74\DeclareMathOperator{\Ext}{Ext}75\DeclareMathOperator{\Res}{Res}76\DeclareMathOperator{\Kl}{Kl}77\DeclareMathOperator{\id}{id}78\DeclareMathOperator{\rank}{rank}79\newcommand{\E}{\mathcal E}80\newcommand{\Spc}{\mathcal S}81\newcommand{\R}{\mathbb R}82\newcommand{\C}{\mathbb C}83\newcommand{\Z}{\mathbb Z}84\newcommand{\overB}{\E_{/B}}85\tikzcdset{every label/.append style={font=\small}}86\begin{document}87\twocolumn[88\begin{@twocolumnfalse}89\noindent\textcolor{paperaccent}{\rule{\textwidth}{1pt}}\par90\vspace{11pt}91{\noindent\fontsize{27}{32}\selectfont\bfseries92Differential Topology of\\Parametrized Neural Networks\par}93\vspace{8pt}94{\noindent\large\color{papermuted}95Jet representations, decision boundaries, and reconstruction\par}96\vspace{10pt}97{\noindent\small November 2026\par}98\vspace{9pt}99\noindent\begin{minipage}{\textwidth}\small100\textbf{Abstract.}101A smooth neural architecture represents a space of weights by evaluations102on a data manifold. We give architecture-level conditions for controlling103first jets, decision boundaries, and reconstruction. An affine readout of104immersed features is submersive on first jets, while an arbitrary smooth105feature map with free final biases still gives generic transverse106classification boundaries. We compute a fold discriminant for a107quadratic readout on the circle and a regular family whose decision108boundaries have nontrivial monodromy. For equivariant layers the109stabilizer representation can force an excess intersection even when110the total evaluation is submersive. An explicit torus calculation111exhibits the associated obstruction complex. For autoencoders, exact112ambient recovery requires an embedding, and a spherical example gives113a sharp uniform error bound below the required latent dimension.114Relative mapping objects and descent specify how these results apply115to local parameter and feature presentations.116\par\medskip117\noindent\textbf{Keywords.}\enspace118Parametrized networks, transversality, representation theory, singularities,119monodromy, autoencoders.120\end{minipage}121\vspace{17pt}122\end{@twocolumnfalse}123]124\thispagestyle{plain}125 126\section{Networks as families of smooth maps}127\label{sec:families}128 129A parametrized neural network is a family of maps realized by a130specified architecture. Differential topology studies these maps131through their derivatives, their fibers, and their behavior along132paths of parameters. This makes it possible to ask which local133perturbations a layer actually supplies and which global topological134constraints remain even if local perturbations are plentiful.135A basic example is an encoder with a smooth, trainable readout.136The readout can have enough derivatives to move a decision boundary,137while a narrow intermediate space still obstructs reconstruction138on the whole data manifold.139 140The starting points are Hirsch's treatment of jets and transversality