Indexed text attachment
differential-topology-of-parametrized-neural-networks.tex
Source for Differential Topology of Parametrized Neural Networks
View paper · Download original file
Lines 214–313 of 1074. Source text is displayed literally.
214use the affine readout215\begin{equation}216 \label{eq:affine-readout}217 f_{A,b}(x)=A\phi(x)+b,\qquad218 (A,b)\in\mathcal A219 =\operatorname{Hom}(\mathbb R^d,\mathbb R^q)\times\mathbb R^q .220\end{equation}221The choice to hold the feature extractor fixed specifies a subfamily222of whatever full architecture realizes $\phi$. This separates the223derivatives supplied by the readout from those provided by224perturbing upstream layers.225 226\begin{theorem}[First-jet control by an affine readout]227\label{thm:jet-surjectivity}228If $\phi$ is an immersion, then229$J:\mathcal A\times M\to J^1(M,\mathbb R^q)$,230$(A,b,x)\mapsto j^1_x f_{A,b}$, is a submersion.231Without the immersion assumption, the corresponding map to232$J^0(M,\mathbb R^q)$ remains a submersion. In either case, for a233finite or countable family of smooth jet submanifolds at the234indicated order, the parameters for which the corresponding235jet section is transverse to all of them form a conull residual236set.237\end{theorem}238 239\begin{proof}240At fixed $x$, use jet coordinates $(y,L)$ with241$L:T_xM\to\mathbb R^q$. The differential in the readout242directions is243\[244 (\dot A,\dot b)245 \longmapsto246 \bigl(\dot A\phi(x)+\dot b,\ \dot A D\phi_x\bigr).247\]248Given any $L_0:T_xM\to\mathbb R^q$ and immersive $\phi$, extend249$L_0(D\phi_x)^{-1}$ linearly from $D\phi_x(T_xM)$ to250$\mathbb R^d$ to obtain $\dot A$. Then any value variation251$y_0$ is realized by $\dot b=y_0-\dot A\phi(x)$.252This proves surjectivity in the vertical jet coordinates.253Base coordinates are covered because $J$ lies over the254projection to $M$. For $J^0$ only the value variation is255needed, and free biases supply it at every $x$.256 257A submersion is transverse to each smooth jet submanifold $S$.258Its preimage is smooth, and regular values of the projection259of the preimage to $\mathcal A$ are exactly the parameters260whose jet extensions are transverse to $S$. This is the261parametric transversality argument262\cite[Ch.~6]{higher-hirsch}. Sard's theorem gives a null263exceptional set. The critical locus has a countable compact264exhaustion because it is closed in a second-countable265manifold. Its images under the projection are compact266null sets and therefore nowhere dense. Taking the countable267intersection of the good parameter sets proves the claim.268\end{proof}269 270For a binary classifier, Theorem~\ref{thm:jet-surjectivity}271implies that generically the decision boundary is empty or a272codimension-one smooth submanifold, even if $\phi$ has273singular points. The free final bias alone suffices. A274fixed-bias family can have additional constraints.275For example, if a feature identically vanishes on a276region, a bias-free readout may vanish on that entire region.277For vector outputs, immersion of $\phi$ supplies every278first-order normal perturbation at a given input. It makes279no assertion that independent perturbations at multiple280inputs can be realized by the same $\dot A$.281 282\subsection{Several classes and several points}283 284Consider $q$ logits and the open part of the tie stratum285where a subset $I$ of $r$ classes is exactly the set of286maximizers. Fix $i_0\in I$. Equalities287$y_i-y_{i_0}=0$ for $i\in I\setminus\{i_0\}$ are $r-1$288independent linear equations, and strict inequalities289separate the other coordinates. This is a smooth290codimension-$(r-1)$ submanifold of $\mathbb R^q$.291Consequently, the inverse image for a generic readout292is empty or has codimension $r-1$. No $r$-way maximum293tie occurs if $r>n+1$. The conclusion also holds when294the bias space is restricted to vectors with coordinate295sum zero. The differences of any bias vector are296unchanged by subtracting a constant vector so the297variation still surjects onto the normal directions to298each tie. Thus the ordinary redundancy of adding the299same constant to every logit can be removed without300losing the relevant perturbations.301 302There is a complementary global result when $\phi$303is an embedding. For different points $x,x'$ the304map $(A,x,x')\mapsto A(\phi(x)-\phi(x'))$ is a submersion305to $\mathbb R^q$. Indeed, $\phi(x)-\phi(x')$ is nonzero306and a linear map can assign it any value. The source307in the input variables is $M\times M\setminus\Delta$,308of dimension $2n$. When $q>2n$, parametric309transversality forces absence of off-diagonal coincidences310for a conull residual set of $A$. By the first-jet argument,311$r$-rank derivative strata of codimension312$(n-r)(q-r)$ are avoided when that codimension exceeds $n$.313This formula follows by choosing an invertible $r$-minor