Indexed text attachment
topos-actions-in-machine-learning-and-ai.tex
Source for Topos Actions in Machine Learning and Artificial Intelligence
View paper · Download original file
Lines 901–1000 of 1207. Source text is displayed literally.
901 902Let the deployment base be $B=S^1=\mathbb R/(2\pi\mathbb Z)$.903In $\operatorname{Sh}(B)$ let $G$ be the locally constant904group sheaf with fiber $C_3=\mathbb Z/3$ and monodromy905$j\mapsto-j$. Only $0$ is fixed by inversion, so906$\Gamma(B,G)$ is the trivial group. The local groups are907nevertheless nontrivial.908 909Use real regular-representation feature fibers910$V_t=\mathbb R[C_3]$ with basis $(e_j)_{j\in C_3}$.911The monodromy on the bundle $V$ is the orthogonal912permutation $S e_j=e_{-j}$. Let $R e_j=e_{j+1}$913be the group shift in a local lift of the angular914coordinate. Then $SRS^{-1}=R^{-1}$, which makes the915regular action compatible with the group sheaf's916transition. The sheaf of smooth sections of $V$917is an internal $G$-object. Its fiberwise $G$-linear918maps define equivariant linear layers in the relative919classifying topos.920 921\begin{proposition}[The coefficient bundle of equivariant filters]922\label{prop:twisted}923The bundle of fiberwise linear intertwiners from $V$924to itself is $\mathbb R^2\oplus L$ over $S^1$, where925$L$ is the real M\"obius line and $\mathbb R^2$ denotes926a trivial rank-two bundle. A flat filter, meaning927locally constant coefficients in the given flat928frames, has two free real coefficients. Smooth929filters can also use the third coefficient as a930section of $L$.931\end{proposition}932 933\begin{proof}934A linear map commuting with $R$ is determined by935its value $w$ on $e_0$, since its values on the other936basis vectors must be $Rw$ and $R^2w$. Every $w$937gives a map by this rule. Hence the commutant consists938precisely of $c_0 I+c_1R+c_2R^2$. Changing frame by939$S$ exchanges $c_1,c_2$. In the basis940$I,R+R^2,R-R^2$ it is the identity on the first941two directions and multiplication by $-1$ on the942third. This gives the bundle splitting and the943flat coefficient count.944\end{proof}945 946The distinction affects model capacity. Using the947global section group alone would impose no commutation948conditions and would allow all nine entries of a local949matrix. Requiring internal equivariance leaves three950local coefficients. Choosing flat global coefficients951ties the two shifted terms together, leaving two.952For smooth models the third coefficient $b(t)$ can953be any smooth function on the lifted angle satisfying954$b(t+2\pi)=-b(t)$. For example955$b(t)=\cos(t/2)$ gives a well-defined global filter956$b(t)(R-R^2)$. Every continuous such $b$ has a zero:957if $b(t)$ is nonzero, the values at $t$ and958$t+2\pi$ have opposite signs. Hence a model that959requires this coefficient to stay nonzero can work over960the orientation double cover, where the pulled-back961line is trivial. Learning this coefficient in the962bundle and learning one untwisted scalar are different963parameterization problems.964 965\begin{figure}[tbp]966\centering967\(968\begin{tikzcd}[column sep=4.5em,row sep=2.2em]969 V_t \arrow[r,"A_t"] \arrow[d,"S"'] & V_t \arrow[d,"S"]\\970 V_{t+2\pi}\arrow[r,"A_{t+2\pi}"'] & V_{t+2\pi}971\end{tikzcd}972\)973\caption{A filter descends when it intertwines the change of974frame. Conjugation by the transition $S$ reverses the975coefficient of $R-R^2$.}976\label{fig:twisted-filter}977\end{figure}978 979A coordinatewise activation commutes with the local shift980and the frame permutation. Consequently one may stack981these linear filters with such activations to obtain a982global nonlinear layer. A sum of the three output983coordinates then gives a scalar readout, invariant under984both the local group and the frame transition.985 986A simple regression calculation gives a training987interpretation with a specified observation law.988In a local frame draw $x=e_j$ uniformly over $j$,989draw a hidden sign $\varepsilon\in\{\pm1\}$ fairly990and independently, and observe the target991$y=R^\varepsilon x$. Inverting the frame sends992$(x,y,\varepsilon)$ to $(Sx,Sy,-\varepsilon)$, so993the law descends even though a deterministic forward994shift $R$ does not define a flat global filter.995The least-squares estimator conditional on $x$ is996$(R+R^2)x/2$, an admissible flat filter.997The two conditional target vectors are distinct basis998vectors, so each lies squared distance $1/2$ from999their mean. The best expected squared error using1000only $x$ is therefore $1/2$. The rule $(x,\varepsilon)\mapsto R^\varepsilon x$ has zero error