Organizations: Institute for Mathematics, Image and Pattern Analysis Group, Heidelberg University, Germany · Research Station Geometry and Dynamics, Heidelberg University
This paper introduces a geometric approach to the representation and learning of time-variant parametrizations of deep assignment flows, for binary data labelings on graphs. Each edge carries an optimally controlled parameter flow in a 2-dimensional totally geodesic submanifold, based on objectives derived from a non-local linear scale space representation of the input data. The aggregation of triplets of such flows in 5-dimensional totally geodesic submanifolds is learned from data. The fusion of these flows determines the trajectory in the high-dimensional symmetric parameter space and by geometric integration the layers of binary deep assignment flows. Numerical results demonstrate the expressivity and the limitation of our highly parallel geometric local synthesis approach to high-dimensional deep assignment flow parametrizations.
Figures & tables
Figure 1.1 : Illustration of our geometric local-to-global synthesis approach of high-dimensional parameter trajectories ( 1.1g ) for deep binary assignment flows, via the optimal control of trajectories and their aggregation in totally-geodesic submanifolds P12,P13 .
Figure 3.1 : Left : Illustration of the layer mapping and its parametrization of the local binary assignment network ( 2.3b ), for an arbitrary edge ij∈E(G) of the underlying graph (dropped above from the notation) and the corresponding pairs of input feature vectors and output assignment vectors. The control parameters ( 3.2 ) which determine the parameter matrix, determine in turn the flow. We learn local binary assignment flows by determining optimal trajectories of these parameters by discrete optimal control. Right : The feasible set for vectors of the sequences ( 3.3b ). Regions near the origin corresponding to a large uncertainty of local label assignments and hence are discretized more finely, whereas regions farther from the origin are discretized more coarsely.
Figure 3.2 : Left: A triangle (minimally complete) subgraph. Each edge indexes an optimally controlled binary assignment flow network ( 2.3b ). Fusing these minimal networks into a larger network creates conflicts because each controlled variable participates in two local systems, as indicated by the above right-most panel and the application of the injection map ( 3.10 ) to the three sequences ( 3.9 ). This problem is resolved by a machine learning component (center panel) with learnable parameters θte,e∈{ij,ik,jk} .
Figure 3.3 : Performance of locally labeled triangular subgraphs, based on local geometric averaging ( 3.13 ) of three 2×2 flows indexed by the edges. Each point indicates an initial point around 0 , mapped to a small sphere for visualization. Its color displays if the trajectory ( 2.3b ) reaches the target octant (−−+) (green) or makes a single (blue), two (black) or three (red) labeling errors at the vertices. Errors occur when the starting and target octants differ by ≥2 binary labels.
Figure 3.4 : Image labeling via a linear scale space, exemplified for a corrupted handwritten digit “8”. Top row: Scale-space representations v1↦G(⋅,τd)∗v1,d∈[D] , of the input data. Bottom row: In the training phase, analogous representations of the target labeling vT+1∗↦G(⋅,τd)∗vT+1∗ are provided as well. Center rows, inference: The learned sequence of parameter matrices Ωt,t∈[T] enables the binary assignment flow network ( 2.3 ) to infer the uncorrupted labelings (3rd row; green arrows) from the corrupted labelings (2nd row; yellow arrows) across the scale space.
Figure 4.1 : Left : Examples of input images (top) corrupted initial labelings (mid) and target labelings (bottom) from the Rotated MNIST data base. Right : Examples of generating images, initial labelings and target labelings for the Dino scene.
Figure 4.2 : Triangular subgraphs. Rotated MNIST (leftmost 2 panels): Vertices associated with a center vertex (marked red, left panel), used to define triangular subgraphs (right panel, different vertex embedding for visualization). Dino scene (rightmost 2 panels): Vertices associated with a center vertex (marked red, left panel), used to define triangular subgraphs (right panel, different vertex embedding for visualization). For both scenes, every vertex becomes a center vertex (red) and thus defines a local collection of triangular subgraphs.
Rotated MNIST
Diff. Stage
Foreground Acc.
Background Acc.
0
97.26 %
98.83 %
1
96.67 %
98.62 %
2
92.83 %
98.34 %
Table 1: Prediction performance for independent test data for Rotated MNIST (left) and the Dino scene (right). Recall from Fig. 3.4 that during inference, the sequence of diffusion stages is 2→1→0 . The accuracy measurements have been averaged over the entire test data set.
Figure 4.3 : Predicted labelings for independent test data for Rotated MNIST (left) and the Dino scene (right). Pink pixels indicate vertices with wrong label assignments.
We investigate the dependence of the approximation capacity of deep residual networks on its depth in a continuous dynamical systems setting. This can be formulated as the general problem of quantifying the minimal time-horizon required to approximate a diffeomorphism by flows driven by a given family F of vector fields. We show that this minimal time can be identified as a geodesic distance on a sub-Finsler manifold of diffeomorphisms, where the local geometry is characterised by a variational principle involving F. This connects the learning efficiency of target relationships to their compatibility with the learning architectural choice. Further, the results suggest that the key approximation mechanism in deep learning, namely the approximation of functions by composition or dynamics, differs in a fundamental way from linear approximation theory, where linear spaces and norm-based rate estimates are replaced by manifolds and geodesic distances.
Jingpu Cheng, Qianxiao Li, Ting Lin +1
Department of Mathematics, National University of Singapore, 117543, Singapore · Department of Mathematics and Institute for Functional Intelligent Materials, National University of Singapore, 117543, Singapore · School of Mathematical Sciences, Peking University, 100871, China
Recent work has shown that models flow matching models can be trained without explicit time conditioning, challenging the standard view that the interpolation time is needed to disambiguate velocity targets. But why should a time-blind model work at all? Decomposing the time-blind flow matching loss, we identify two sources of irreducible error: a coupling variance, which arises from ambiguous velocity targets induced by how noise and data points are paired, and the time-blindness gap, which is the additional error caused by ignoring time. This gap shows that time-blind training is strictly harder than conventional training, reinforcing the puzzle that time-blind models work so well in practice. We resolve this tension by showing that the geometry of high-dimensional data makes time identifiable directly from noisy observations. When data concentrates near a k-dimensional subspace, time can be recovered from the statistical structure of noisy interpolants in directions orthogonal to the data; under a spiked-covariance model, this yields a closed-form estimator that recovers t from a single observation z at rate O(1/d−k) for ambient dimension d. As a consequence, we prove that the time-blindness gap is asymptotically negligible relative to the coupling variance. We empirically demonstrate our identifiability result on real-world data and show that changing the coupling has a much larger effect on loss and sample quality than removing time conditioning across CIFAR-10, CelebA-HQ, and FFHQ. These results explain why time-blind flow matching works and show that the main practical lever is the choice of coupling, not explicit time conditioning.
Alec Helbling, Sebastian Gutierrez Hernandez, Benjamin Hoover +2
Riemannian manifolds are powerful models for problems with non-Euclidean structure. Recent work in geometric machine learning has focused on training deep neural networks to learn Riemannian metrics from high-dimensional data. However, many geometric algorithms require repeated evaluations of the exponential and logarithmic maps, whose high computational cost restricts the utility of learned Riemannian manifolds in practice. We propose to address this cost by training a conditional invertible neural network, called a flow chart, to approximate normal coordinates around points on a given Riemannian manifold. With this approach, the exponential and logarithmic maps can be approximated by fast Euclidean operations in the model's learned coordinate space. Our main contribution is an efficient loss function for training flow charts that is nonnegative and, under suitable domain assumptions, is zero precisely when the chart is normal. Across sixteen synthetic and data-derived geometries, flow charts are competitive with a specialized neural geodesic interpolator for logarithmic map estimation, while also being able to estimate the exponential map. When pretrained, flow charts provide a median speedup of 60-70x over numerical baselines at comparable logarithmic map estimation error, and their learned coordinates improve the accuracy of a numerical geodesic solver.