Received: 2 September 2024 Accepted: 30 September 2024

DOI: 10.1002/aaai.12210

H IGH LIGH T

Geometric Machine Learning

Melanie Weber

School of Engineering and Applied

Sciences, Harvard University, Cambridge, Abstract
Massachusetts, USA A cornerstone of machine learning is the identification and exploitation of struc-
Correspondence
ture in high-dimensional data. While classical approaches assume that data lies
Melanie Weber, School of Engineering in a high-dimensional Euclidean space, geometric machine learning methods are
and Applied Sciences, Harvard University, designed for non-Euclidean data, including graphs, strings, and matrices, or data
Cambridge, Massachusetts, USA.
Email: mweber@[Link] characterized by symmetries inherent in the underlying system. In this article,
we review geometric approaches for uncovering and leveraging structure in data
Funding information
and how an understanding of data geometry can lead to the development of more
NSF, Grant/Award Number:
CBET-2112085,DMS-2406905 effective machine learning algorithms with provable guarantees.

INTRODUCTION et al. 2024), biochemistry (Azizzadenesheli et al. 2024;

Diepeveen et al. 2024; Ganea et al. 2021; Watson et al.
Many classical machine learning methods assume that 2023; Zhao and Singer 2014), climate science (Kashinath
data lies in a high-dimensional Euclidean space. How- et al. 2021; Li et al. 2021), biomedical sciences (Bekkers
ever, many applications involve structured data, such et al. 2018; Bhasker et al. 2024; Klimovskaia et al. 2020;
as sets, graphs, and matrices, data concentrated in low- Weber et al. 2017c), and robotics (Wang et al. 2022), among
dimensional subspaces, and data exhibiting symmetries others. The development of geometric machine learning
derived from fundamental laws of physics. Recognizing approaches is guided by two central questions:
and leveraging such structure can lead to more efficient
and interpretable models. 1. How can we characterize useful geometric structures in
One of the early examples of this idea are convolutional data?
neural networks (CNN) (Krizhevsky, Sutskever, and 2. How can we design (provably) more efficient algo-
Hinton 2012; LeCareun et al. 1998), which revolutionized rithms and architectures that explicitly encode such
image classification in the 2010s. In image classification, structure?
the goal is to assign labels based on the objects within
an image, regardless of their position. CNNs achieve The first question relates to data geometry, characterizing
this by encoding translation invariance as an inductive which involves understanding the local and global geo-
bias, that is, by restricting the model to functions whose metric properties of data on both discrete and continuous
output is unchanged under translations of the input. domains. The second question concerns the design of
Geometric machine learning extends this idea by encod- algorithms and architectures that leverage these geometric
ing various types of geometric structures into model properties and the fundamental question of when and why
architectures (Bronstein et al. 2021; Cohen and Welling geometric approaches may outperform classical methods.
2016). These approaches have been particularly impactful This article discusses key geometric tools for character-
in scientific machine learning, with applications spanning izing data geometry and how such characterizations can
material sciences (Batzner et al. 2022; Subramanian be leveraged in learning tasks on graphs, matrix spaces,

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium,
provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2025 The Author(s). AI Magazine published by John Wiley & Sons Ltd on behalf of Association for the Advancement of Artificial Intelligence.

AI Magazine. 2025;46:e12210. [Link]/journal/aaai 1 of 11

[Link]
2 of 11 AI MAGAZINE

and symmetric spaces. It accompanies a New Faculty reach can ensure that the manifold does not have sharp
Highlights talk given at AAAI 2024 (Weber 2024). corners or other irregularities. Another concept that is use-
ful in this context is Ricci curvature, a local, intrinsic notion
of curvature, which describes how the volume within small
CHARACTERIZING DATA GEOMETRY regions of the manifold changes.

The first step in leveraging data geometry is to identify

and parametrize such structure in a way that is amenable Discrete setting
to downstream applications. Data and models can exhibit
various geometric structures, such as symmetries aris- As discussed above, graph- and set-structured data, which
ing from fundamental physical laws or low-dimensional are prime examples of discrete data structures, are inher-
structure that arises from a few latent factors. ently permutation-invariant. As such, their structure can
be characterized through the lens of permutation sym-
metries. Another approach to studying their structure
Continuous setting involves discrete notions of curvature. This requires the
extension of classical concepts of curvature, which are typ-
Symmetries ically defined in continuous settings, to discrete spaces.
Symmetries, sometimes called “invariances,” are trans- We introduce such notions below and describe how they
formations that leave crucial properties of the input data can be used to analyze both local and global geometric
unchanged (or “invariant”). Mathematically, such struc- properties of data.
ture can be characterized with algebraic and geometric
tools through the language of group theory: The set of Discrete Notions of Curvature
transformations 𝑈 ∶  → , which, when applied Curvature is a fundamental concept in Differential Geom-
to a data point 𝑥 ∈ , do not change the value of the etry, which characterizes the geometric properties of
objective (i.e., 𝑓(𝑈𝑥𝑈 † ) = 𝑓(𝑥) for all 𝑥 ∈ ) has a group geodesic spaces. Here we consider discretizations of Ricci
structure 𝐺. This concept is crucial in understanding curvature that allow for characterizing the local and global
a variety of structures in machine learning and data geometric properties of data represented in discrete spaces,
science applications. For example, in relational data like such as graphs, networks, hypergraphs, and simplicial
sets or graphs, there is often no natural order relation complexes. One of the key challenges in defining curva-
among the elements, which is indicative of permutation ture in discrete spaces is the lack of a (natural) differential
symmetries. In image classification, the focus is usually structure. To address this, we define curvature using cur-
on identifying the type of object in the image, regardless of vature analogies: In continuous spaces, the curvature is
its location, which corresponds to translation symmetry. related to several classical tools, many of which have well-
When predicting the structural properties of molecules, studied counterparts in discrete settings. By defining a
the output should not be affected by global rotations curvature concept that maintains the relationship with
of the input, that is, it should be rotation-invariant. In one of these discrete characteristics, we can define a dis-
these cases, it is important that machine learning models crete curvature notion that retains some properties of its
respect these fundamental structural properties of the continuous counterpart.
data. We will later discuss how to encode such structural Two curvature notions, introduced by Ollivier (Ollivier
information as inductive bias into machine learning 2010) and Forman (Forman 2003), are particularly pop-
architectures. ular in machine learning on graph-structured data.
Ollivier’s curvature is based on a key relationship between
Smooth Data Manifolds Ricci curvature and the behavior of random walks.
In learning tasks it is often assumed that high-dimensional We will introduce this notion in detail below and dis-
data lies on or near a lower-dimensional manifold (the so- cuss how it motivates the applications discussed in the
called “manifold hypothesis”). Leveraging such structure next section. Forman’s curvature, while closely related
in downstream tasks requires characterizing the geometric to Ollivier’s, is often used as a more scalable alter-
properties of this manifold, such as its intrinsic dimension native. Although Ollivier’s curvature typically offers a
and curvature. It is common to assume certain smoothness richer geometric characterization, Forman’s curvature
conditions, for instance, by bounding the reach. The reach is computationally more efficient, especially for large
of a manifold corresponds to the largest distance at which and dense graphs. Consequently, applications of both
points in the surrounding space still have a unique closest notions have been explored in the machine learning
point on the manifold (Federer 1959). Restrictions on the literature.
AI MAGAZINE 3 of 11

uous regimes. A common setting in which such limits are

studied are geometric graphs that are constructed from 𝜖-
nets of a point cloud (i.e., nodes corresponding to pairs of
points are connected, if they are within distance 𝜖) and cap-
ture spatial information in the underlying data sets. Here,
continuum limits characterize the geometry of the graph
as the number of nodes grows larger and the resolution
becomes finer. Limits of graph characteristics have found
applications in the study of point clouds, in geometry
processing tasks within computer vision, as well as in man-
ifold learning, a classical task in geometric data analysis.
FIGURE 1 Computing Ollivier’s Ricci curvature.
Continuum limits have been explored for a range of
geometric characteristics. A key example are geodesic
Ollivier’s Ricci Curvature distances, where measuring distances in the (discrete)
Ollivier (Ollivier 2010) introduces a Ricci curvature, which observation space can allow for inferring the “true” dis-
relates the curvature along a geodesic between nearby tances on the underlying data manifold (see, e.g., Davis and
points 𝑥, 𝑦 on a manifold  with the optimal transporta- Sethuraman; Diaz et al. 2016). Another important example
tion distance between their neighborhoods. is the Graph Laplacian and its relation to the Laplace-
This relies on a close relation between the behavior Beltrami operator of the underlying data manifold (Calder
of random walks starting at nearby points, which can and Trillos 2022; Hein, Audibert, and von Luxburg 2005).
be characterized via optimal transport, and Ricci curva- The Graph Laplacian is a tool widely used in graph
ture. Specifically, random walks are likely to draw closer machine learning applications, such as spectral clustering
together if the Ricci curvature is positive and further apart, and semi-supervised learning (Von Luxburg 2007).
if the Ricci curvature is negative. Random walks are well- Recently, continuum limits of discrete Ricci curvature
studied on graphs as well, and one can use this knowledge have also been studied. The first asymptotic convergence
to define a meaningful curvature notion. In particular, guarantee of Ollivier’s curvature was given by (van der
if we start uniform random walks at neighboring nodes Hoorn et al. 2021). More comprehensive non-asymptotic
𝑣1 , 𝑣2 , they are likely to stay close-by, if their neighbor- convergence guarantees, which provide concrete error
hoods overlap significantly, and likely to draw apart if there bounds with respect to the size of the data set, have been
is no significant overlap. This can be quantified using the established in (Trillos and Weber 2023). Importantly, (Tril-
Wasserstein-1 distance 𝑊1 (𝜇1 , 𝜇2 ) between two probabil- los and Weber 2023) also demonstrates that Ollivier’s dis-
ity distributions 𝜇1 , 𝜇2 induced by the random walks (see crete Ricci curvature can be used to learn global bounds on
Figure 1). Then the following notion of curvature is posi- the curvature of the underlying data manifold. This could
tive, if the random walks stay close-by, and negative, if they have implications for manifold learning, where intrin-
draw apart (𝑑𝐺 (𝑥, 𝑦) denoting the shortest path distance): sic curvature offers complementary insights to intrinsic
dimension While intrinsic dimension is implicitly learned
𝑊1 (𝑚1 , 𝑚2 )
𝜅(𝑣1 , 𝑣2 ) = 1 − . by classical algorithms such as MDS and Isomap, it does
𝑑𝐺 (𝑣1 , 𝑣2 ) not allow for a characterization of the data’s intrinsic
curvature. Notably, these non-asymptotic results do not
To compute Ollivier’s curvature in practice, we need to
assume access to the true geodesic distances, a commonly
solve an optimal transport problem for each pair of con-
used assumption that simplifies the analysis, but renders
nected nodes in the graph. This can be done, e.g., using
the results impractical as it requires knowledge of the true
the Hungarian algorithm (Kuhn 1955). Unfortunately, this
data manifold.
can be computationally expensive, especially for large and/
or dense graphs. Faster approximations can be obtained
by replacing the Wasserstein-1 distance with the Sinkhorn
LEARNING ON GRAPHS
distance (Sinkhorn and Knopp 1967) or by using a com-
binatorial approximation of Ollivier’s curvature (Tian,
In this section, we review graph machine learning
Lubberts, and Weber 2023b).
approaches that leverage the inherent geometric structure
of graphs. This includes exploiting permutation symme-
Continuum limits tries in graph- and set-structured data, as well character-
izations of a graph’s geometry via discrete curvature.
Continuum limits of graphs and their geometric character- We briefly comment on another geometric tool, the
istics can provide a bridge between the discrete and contin- Graph Laplacian, which is frequently used in graph
4 of 11 AI MAGAZINE

ciated Ricci flow can be leveraged to refine this detection

procedure. While these methods have primarily focused
on single-membership community detection, where each
node belongs to only one community, recent work has
generalized curvature-based methods to the more difficult
task of mixed-membership community detection (Tian,
Lubberts, and Weber 2023a, 2023b). In these cases,
curvature-based techniques are applied to the line graph,
the dual representation of the original graph, which allows
for better characterizing the geometric structure induced
by nodes that belong to multiple communities.

FIGURE 2 Curvature characterization of a graph. Curvature-Based Coarsening

Graph coarsening allows for reducing the size of an input
graph while preserving its key properties. Typically, this
machine learning, including for clustering and coarsen- involves merging nodes through a process known as edge
ing tasks. Its spectrum is particularly effective at capturing contraction, resulting in a smaller, simplified graph. Sev-
global and mesoscale structures in the graph. In contrast, eral geometric approaches have been leveraged to identify
discrete curvature characterizes the geometric properties candidate edges for contraction, including spectral meth-
of the graph more locally. This makes both approaches, ods (Fiedler 1973; Spielman and Teng 1996), the Louvain
to some extent, complementary. Since computing curva- algorithm (Blondel et al. 2008), and Graclus (Dhillon,
ture only requires access to local information, it can be Guan, and Kulis 2007). From the perspective of discrete
easily parallelized, which enhances scalability, specifically curvature, edges with high curvature are good candidates
on large and dense graphs. On the other hand, computing for contraction, as they represent less critical connec-
the spectrum of the Graph Laplacian requires a spec- tions in the graph. As such, curvature-based pooling
tral decomposition of the adjacency matrix, which can operators, which implement similar algorithmic ideas as
come at a higher computational cost. By understanding presented above in the context of community detection
the strengths and limitations of both tools, practitioners may be leveraged (Feng and Weber 2024; Weber, Saucan,
can better choose the appropriate geometric framework and Jost 2017a, 2017b). Graph Coarsening has applica-
depending on the specific learning task. tions in exploratory data analysis via visualization of the
coarsened graph, as well as the reduction and simplifica-
tion of large input graphs to enable more computationally
Shallow graph machine learning expensive downstream analysis (Weber et al. 2017c).

Curvature-Based Clustering
Community detection, or unsupervised node clustering, is Deep graph machine learning
a fundamental task in graph learning, which seeks to iden-
tify densely connected substructures within a graph. Many, Message-passing Graph Neural Networks
by now classical, algorithms have been proposed for this Graph Neural Networks (GNNs) have emerged as a popu-
task, including spectral clustering (Von Luxburg 2007) and lar architecture for deep learning on graphs. Many GNNs
the Louvain algorithm (Blondel et al. 2008). More recently, implement the message-passing paradigm (Gori, Monfar-
approaches that leverage discrete Ricci curvatures and dini, and Scarselli 2005; Hamilton, Ying, and Leskovec
associated geometric flows have been proposed, including 2017), which iteratively learns node embeddings ℎ𝑣𝑙 that
curvature-based thresholding (Fesser et al. 2023; Gosztolai jointly represent information encoded in the adjacency
and Arnaudon 2021; Sia, Jonckheere, and Bogdan 2019), as and the attributes of the input graph (Figure 3). During
well as reweighting and thresholding guided by an associ- each iteration (“layer”) 𝑙, nodes aggregate information
ated discrete Ricci flow (Ni et al. 2019; Tian, Lubberts, and from their neighbors (the “message” 𝑚𝑣𝑙 ) and then update
Weber 2023b). Both types of methods rely on the obser- their own representation based on this new information.
vation that the curvature of edges between communities This allows for capturing increasingly complex structure
is low, whereas edges within communities have high cur- as the depth of the network increases. The aggregation
vature (Figure 2). Curvature-based thresholding removes 𝑓𝐴𝑔𝑔 and update functions 𝑓𝑈𝑝 that specify each layer
edges with low curvature, upon which the resulting are typically implemented using neural networks with
connected components reveal the communities. The asso- trainable parameters. Examples of message-passing GNNs
AI MAGAZINE 5 of 11

include Graph Convolutional Networks (Kipf and Welling suitable number of edges to add and remove. Here, again,
2017), Graph Isomorphism Networks (Xu et al. 2018), the geometry of the graph can provide guidance: The cur-
GraphSAGE (Hamilton, Ying, and Leskovec 2017), and vature gap (Gosztolai and Arnaudon 2021) is a global graph
Graph Attention Networks (Veličković et al. 2017). characteristic derived from the distribution of curvature
values across the graph, from which a threshold for identi-
Challenges during training fying regions with “too high” or “too low” curvature can be
While Graph Neural Networks have achieved significant deduced. This, in turn, informs the number of edges that
success across various learning tasks and domains, they are need to be perturbed for effective rewiring.
not without challenges. Two primary issues are the inabil-
ity to effectively utilize information encoded in long-range Improving the representational power of GNNs via
connections (an effect known as over-squashing (Alon and geometric augmentation
Yahav 2021)), and the difficulty in maintaining distinct Another key question regarding the effectiveness of GNNs
node representations as the network depth increases is what types of functions they can and cannot learn.
(known as over-smoothing (Li, Han, and Wu 2018)). This question can be analyzed through the following lens:
Over-squashing occurs when information encoded in Does a GNN map two graphs to the same representation if
long-range connections is compressed through “bottle- and only if they are topologically identical? Unfortunately,
necks” in message-passing, which can make it difficult to standard message-passing GNNs are unable to learn such
capture this information in the learned representations as a mapping for notable classes of graphs, including simple
the number of layers in the network grows. From a geo- examples such as regular graphs (Morris et al. 2019; Xu
metric perspective, it can be observed that edges that cause et al. 2018). As a consequence, GNNs are unable to learn
such “bottlenecks” have low curvature (Topping et al. some basic functions, such as calculating the diameter
2022). On the other hand, over-smoothing happens when, of a graph or determining the size of its largest cycle.
with increasing network depth, the representations of dis- What can be done to address this? More powerful GNN
similar nodes become increasingly indistinguishable. This architectures, such as Higher-order GNNs (Morris et al.
arises in particular in dense regions of the graph, which 2019) and topology-aware message passing that encodes
are often characterized by high curvature (see Figure 2). local substructures (Bouritsas et al. 2022; Zhao et al. 2022)
As such, discrete curvature offers a useful framework are able to distinguish a larger set of non-topologically
for characterizing both effects. Curvature can also be identical graphs. However, often these architectures are
used to mitigate both effects during training, specifically computationally less efficient. A simpler alternative is
through rewiring, a preprocessing technique that perturbs to augment the graph with geometric information that
the edges of the input graph. Several rewiring approaches allows for learning richer node representations, which
have been proposed, many guided by graph characteris- enhances their utility in downstream tasks. Various types
tics like discrete curvature (Topping et al. 2022; Fesser of such augmentations, also known as “encodings”, have
and Weber 2023; Nguyen et al. 2023) or the spectrum of been introduced, leveraging tools like the spectrum of
the Graph Laplacian (Karhadkar, Banerjee, and Montufar the Graph Laplacian (Dwivedi et al. 2022) and discrete
2023). Curvature-based rewiring performs the following curvature (Fesser and Weber 2024).
edge perturbations: In regions with high curvature, edges As we discussed above, curvature captures local struc-
are removed to prevent over-smoothing, while in areas tural information within a node’s two-hop neighbor-
with low curvature, synthetic edges are added to coun- hood in contrast to message-passing, which only cap-
teract over-squashing. A key question is to determine a tures information in the one-hop neighborhood. As a

FIGURE 3 Message-passing paradigm.

6 of 11 AI MAGAZINE

consequence, simple curvature augmentations are surpris- tasks is of great interest. In the Euclidean setting, Dis-
ingly effective encodings and can lead to significant per- ciplined Convex Programming (DCP) (Grant, Boyd, and
formance improvements, often outperforming encodings Ye 2006) is a well-established framework that automates
based on other graph characteristics (Fesser and Weber the verification of convexity for a wide range of func-
2024). Importantly, curvature encodings are relatively tions. DCP achieves this by decomposing functions into
scalable compared to other methods. Since encodings rep- basic convex components (atoms) and applying convexity-
resent a preprocessing routine, scalability is important to preserving operations (rules). This idea can be extended to
balance the benefits of enhanced representational power the geometric setting by leveraging the algebraic properties
with the computational resources required for computing of g-convex functions. Disciplined Geodesically Convex Pro-
the encodings. gramming (Cheng, Dixit, and Weber 2024) generalizes this
framework to Cartan-Hadamard manifolds, specifically
for optimization on the manifold of symmetric, positive
LEARNING ON MATRIX SPACES definite matrices.

In many machine learning and data science applications Algorithms for Riemannian optimization
we encounter optimization tasks that involve matrices, Motivated by the advantages discussed above, Riemannian
both in the form of standalone algorithms or as subroutines optimization has gained significant traction in machine
of more complex algorithms. Traditionally, these tasks learning and data science. Many classical algorithms have
are solved with tools from Euclidean optimization, which been extended to the geometric setting, resulting in gen-
treat the underlying matrix space as an Euclidean space. eralized approaches for convex (Bacák 2014; Udriste 1994;
Here, the structure of the matrices, such as orthogonality Zhang and Sra 2016), nonconvex (Boumal, Absil, and Car-
or positive definiteness, is enforced through constraints, tis 2019; Weber and Sra 2021), stochastic (Bonnabel 2013;
necessitating the use of constrained optimization tech- Weber and Sra 2021; Zhang, Reddi, and Sra 2016), con-
niques. However, matrix spaces can also be viewed as strained (Bergmann et al. 2022; Weber and Sra 2021, 2022),
Riemannian manifolds, where the structural properties and min-max optimization problems (Jordan, Lin, and
are inherently encoded in the domain’s parametrization. Vlatakis-Gkaragkounis 2022; Martínez-Rubio et al. 2023).
This perspective can offer two significant advantages. Moreover, several numerical software packages for solving
First, because the structure is implicitly encoded in the geometric optimization problems have been developed
problem’s formulation, there is no need to impose explicit across different programming languages (Bergmann 2022;
constraints. This transforms the constrained problem Boumal, Mishra, Absil, and Sepulchre 2014; Townsend,
into an unconstrained one, which is often much easier Koep, and Weichwald 2016), allowing for ease of appli-
to solve. Second, the difficulty of solving an optimization cation in practice. However, the computational overhead
problem depends crucially on its convexity. Interestingly, associated with the use of Riemannian tools in algorithms
many matrix-valued optimization problems that are can be non-negligible. This has motivated the investiga-
nonconvex in the Euclidean setting admit a geodesically tion of “mixed approaches,” where Euclidean algorithms
convex formulation in the Riemannian setting. This allows are used to solve g-convex problems, while geometric tools
for the application of (geodesically) convex optimization are employed to certify global optimality. For instance, in
techniques with global optimality guarantees. problems with a difference of convex structure, a purely
Euclidean CCCP algorithm can be used as a solver, and a
Identifying and certifying geodesic convexity Riemannian analysis can then be leveraged to certify its
Geodesic convexity (short g-convexity) is a simple extension global optimality due to g-convexity (Weber and Sra 2023).
of the classical Euclidean convexity concept to the geomet-
ric setting. Formally, we say that a function 𝑓 is g-convex,
if for all 𝑥, 𝑦 ∈ , the inequality 𝑓(𝛾(𝑡)) ≤ (1 − 𝑡)𝑓(𝑥) + LEARNING UNDER SYMMETRY
𝑡𝑓(𝑦) holds for all 𝑡 ∈ [0, 1]. This corresponds to evalu-
ating the usual convexity condition along a geodesic 𝛾 ∶ Lastly, we consider a setting where geometric structure
[0, 1] → , which represents the shortest path between in data and models arises from symmetries. As discussed
the points 𝑥 = 𝛾(0) and 𝑦 = 𝛾(1) on the manifold ; above, symmetries can be understood through transforma-
its shape is determined by the geometry of the mani- tions of the input space that leave the labels unchanged,
fold. Importantly, g-convexity guarantees the existence of that is, 𝑓(𝑈𝑥𝑈 𝑇 ) = 𝑓(𝑥) for all 𝑥 ∈ . Geometric Deep
a global optimum, which can be found with first-order Learning (Bronstein et al. 2021) provides a framework
methods from any initialization. In order to leverage this for explicitly encoding such structure into models, by
useful property, identifying g-convexity in optimization constraining it to learn only functions with the desired
AI MAGAZINE 7 of 11

FIGURE 4 Learning with and without geometric inductive bias.

symmetries. Such geometric inductive bias can improve when compared to their fully-connected counterparts.
the model’s efficiency, including by improving the model’s Such trade-offs have also been explored for other geo-
ability to generalize well. metric methods, such as kernel regression (Tahmasebi
and Jegelka 2023) and robust classification in hyperbolic
Equivariant Neural Networks space (Weber et al. 2020).
Equivariant neural networks are a deep learning archi- Beyond generalization, one can also analyze the com-
tecture, which encode symmetries as inductive biases plexity of learning algorithms designed for symmetric
by adapting the layers to preserve symmetries via equiv- function classes by assessing how effectively neural net-
ariances, that is, ensuring that 𝑓(𝑈𝑥𝑈 𝑇 ) = 𝑈𝑓(𝑥)𝑈 𝑇 works can be trained. Formally, such an analysis can be
(Figure 4). For example, in a classical CNN, which performed in the statistical query (SQ) framework (Kearns
encodes translation invariance, the convolutional (1998)), where hardness is quantified by the number of
layer is translation-equivariant. Nonlinear layers in queries (akin to steps in a learning algorithm) required
these networks use activation functions that preserve to achieve a certain level of accuracy. Gradient descent,
equivariance. Various architectures for equivariant net- the standard method for training deep learning mod-
works have been proposed to handle specific groups. els, is a prime example of this approach. What can we
These include networks designed for rotation-invariant say about the complexity of learning neural networks,
domains (Cohen, Geiger, Köhler, and Welling 2018), geometric or not? Studies on the hardness of learning
permutation-invariant domains like graphs (Kipf and fully-connected neural networks (i.e., without consider-
Welling 2017) and sets (Zaheer et al. 2017), as well as more ing symmetries) have shown that learning even shallow
general groups (Cohen and Welling 2016; Dym and Maron networks on simple data distributions is computationally
2020; Finzi, Welling, and Wilson 2021; Kiani, Fesser, hard (Chen et al. 2022; Diakonikolas et al. 2020). These
and Weber 2024a; Kondor and Trivedi 2018; Villar et al. findings suggest that, in general, it is not feasible to effi-
2021). ciently learn every function that a small neural network
can represent under typical data distributions. This raises
Representation trade-offs an important question: Can incorporating symmetry as
A key question in the context of equivariant neural net- inductive bias help overcome these computational barri-
works and, in fact, more general architectures, is that ers? Unfortunately, no. Imposing symmetry on a function
of representation trade-offs: What do we gain from geo- class does indeed reduce the computational complexity of
metric inductive biases, such as symmetries? Empirically, learning. For example, for finite groups, the reduction is
equivariant neural networks have been shown to offer proportional to the size of the group (Kiani et al. 2024b).
computational benefits in various applications (Wang, However, simply encoding symmetries does not elimi-
Walters, and Yu 2021), but can this be formalized in terms nate the inherent difficulty of training neural networks.
of computational complexity? Overcoming these challenges requires additional geomet-
Most research on this question has focused on the ric assumptions. For instance, recent work (Kiani, Wang,
impact of symmetry and the corresponding inductive and Weber 2024c) has shown that while learning remains
biases on generalization. Notably, (Bietti, Venturi, and hard under input manifolds with bounded curvature, it
Bruna 2021; Mei, Misiakiewicz, and Montanari 2021) estab- becomes feasible when there are additional constraints on
lished theoretical guarantees for improved generalization the volume of the data manifold, such as those arising from
in shallow linear group-convolutional neural networks bounded Ricci curvature.
8 of 11 AI MAGAZINE

CONCLUSIONS tion Problems With Manifold-Valued Constraints.” Journal of

Optimization Theory and Applications 195(2): 596–623.
Bhasker, Nithya, Hattie Chung, Louis Boucherie, Vladislav Kim,
This article surveyed work at the intersection of geometry
Stefanie Speidel, and Melanie Weber. 2024. “Contrastive Poincaré
and machine learning, focusing on characterizing geomet-
Maps for Single-Cell Data Analysis.” In ICLR 2024 Workshop on
ric structure in data and the design of algorithms and Machine Learning for Genomics Explorations. ICLR.
architectures that leverage such structure to learn more Bietti, Alberto, Luca Venturi, and Joan Bruna. 2021. “On the Sample
efficiently. The field of Geometric Machine Learning has Complexity of Learning Under Geometric Stability.” In Advances
gained significant momentum, not least driven by the suc- in Neural Information Processing Systems, edited by M. Ranzato,
cessful application of geometric architectures in analyzing A. Beygelzimer, Y. Dauphin, P. S. Liang, and J. Wortman Vaughan,
scientific data and developing physics-informed models Vol 34. 18673–84. Curran Associates, Inc.
Blondel, Vincent D., Jean-Loup Guillaume, Renaud Lambiotte, and
that can aid scientific discovery. There are still many chal-
Etienne Lefebvre. 2008. “Fast Unfolding of Communities in
lenges and opportunities in this area, such as encoding
Large Networks.” Journal of Statistical Mechanics: Theory and
more diverse geometric structures into models; develop- Experiment 2008(10): P10008.
ing computational techniques that enhance the scalability Bonnabel, Silvere. 2013. “Stochastic Gradient Descent on Rieman-
of geometric models to ensure broader applicability; and nian Manifolds.” IEEE Transactions on Automatic Control 58(9):
gaining a deeper understanding of when and why geomet- 2217–29.
ric architectures provide algorithmic advantages. Boumal, Nicolas, Bamdev Mishra, P-A Absil, and Rodolphe
Sepulchre. 2014. “Manopt, a Matlab Toolbox for Optimization
on Manifolds.” The Journal of Machine Learning Research 15(1):
AC K N OW L E D G M E N T S
1455–59.
This work was supported by NSF awards CBET-2112085 Boumal, Nicolas, Pierre-Antoine Absil, and Coralia Cartis. 2019.
and DMS-2406905. “Global Rates of Convergence for Nonconvex Optimization on
Manifolds.” IMA Journal of Numerical Analysis 39(1): 1–33.
C O N F L I C T O F I N T E R E S T S TAT E M E N T Bouritsas, Giorgos, Fabrizio Frasca, Stefanos Zafeiriou, and Michael
The author declares that there is no conflict. M. Bronstein. 2022. “Improving Graph Neural Network Expressiv-
ity Via Subgraph Isomorphism Counting.” IEEE Transactions on
Pattern Analysis and Machine Intelligence 45(1): 657–68.
ORCID
Bronstein, Michael M., Joan Bruna, Taco Cohen, and Petar
Melanie Weber [Link]
Veličković. 2021. “Geometric Deep Learning: Grids, Groups,
Graphs, Geodesics, and Gauges.” arxiv.2104.13478.
REFERENCES Calder, Jeff and Nicolás García Trillos. 2022. “Improved Spectral
Alon, Uri and Eran Yahav. 2021, “On the Bottleneck of Graph Neu- Convergence Rates for Graph Laplacians on 𝜀-Graphs and k-
ral Networks and Its Practical Implications.” In International NN Graphs.” Applied and Computational Harmonic Analysis 60:
Conference on Learning Representations. 123–75.
Azizzadenesheli, Kamyar, Nikola Kovachki, Zongyi Li, Miguel Liu- Chen, Sitan, Aravind Gollakota, Adam Klivans, and Raghu Meka.
Schiaffini, Jean Kossaifi, and Anima Anandkumar. 2024. “Neural 2022. “Hardness of Noise-Free Learning for Two-Hidden-Layer
Operators for Accelerating Scientific Simulations and Design.” Neural Networks.” Advances in Neural Information Processing
Nature Reviews Physics 6: 320–28. Systems 35: 10709–24.
Bacák, Miroslav. 2014. “Convex Analysis and Optimization in Cheng, Andrew, Vaibhav Dixit, and Melanie Weber. 2024. “Dis-
Hadamard Spaces.” In Convex Analysis and Optimization in ciplined Geodesically Convex Programming.” arXiv preprint
Hadamard Spaces. de Gruyter. arXiv:2407.05261.
Batzner, Simon, Albert Musaelian, Lixin Sun, Mario Geiger, Cohen, Taco, and Max Welling. 2016. “Group Equivariant Con-
Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess volutional Networks.” In International Conference on Machine
E. Smidt, and Boris Kozinsky. 2022. “E(3)-Equivariant Graph Learning, 2990–99. PMLR.
Neural Networks for Data-Efficient and Accurate Interatomic Cohen, Taco S., Mario Geiger, Jonas Köhler, and Max Welling.
Potentials.” Nature Communications 13(1): 1–11. “Spherical CNNs.” arXiv preprint arXiv:1801.10130.
Bekkers, Erik J., Maxime W. Lafarge, Mitko Veta, Koen A. J. Davis, Erik and Sunder Sethuraman. “Approximating Geodesics Via
Eppenhof, Josien P. W. Pluim, and Remco Duits. 2018. “Roto- Random Points.” The Annals of Applied Probability 29(3): 1446–
Translation Covariant Convolutional Networks for Medical Image 86.
Analysis.” In Medical Image Computing and Computer Assisted Dhillon, Inderjit S., Yuqiang Guan, and Brian Kulis. 2007. “Weighted
Intervention–MICCAI 2018: 21st International Conference, Proceed- Graph Cuts Without Eigenvectors a Multilevel Approach.” IEEE
ings, Part I, 440–48. Granada, Spain: Springer. Transactions on Pattern Analysis and Machine Intelligence 29(11):
Bergmann, Ronny. 2022. “[Link]: Optimization on Manifolds in 1944–57.
Julia.” Journal of Open Source Software 7(70): 3866. [Link] Diakonikolas, Ilias, Daniel M. Kane, Vasilis Kontonis, and Nikos
org/10.21105/joss.03866. Zarifis. 2020. “Algorithms and sq Lower Bounds for Pac Learn-
Bergmann, Ronny, Roland Herzog, Julián Ortiz López, and Anton ing One-Hidden-Layer Relu Networks.” In Conference on Learning
Schiela. 2022. “First-and Second-Order Analysis for Optimiza- Theory, 1514–39. PMLR.
AI MAGAZINE 9 of 11

Diaz, Josep, Dieter Mitsche, Guillem Perarnau, and Xavier Pérez- Jordan, Michael, Tianyi Lin, and Emmanouil-Vasileios Vlatakis-
Giménez. 2016. “On the Relation Between Graph Distance and Gkaragkounis. 2022. “First-Order Algorithms for Min-Max Opti-
Euclidean Distance in Random Geometric Graphs.” Advances in mization in Geodesic Metric Spaces.” Advances in Neural Infor-
Applied Probability 48(3): 848–64. mation Processing Systems 35: 6557–74.
Diepeveen, Willem, Carlos Esteve-Yagüe, Jan Lellmann, Ozan Karhadkar, Kedar, Pradeep Kr. Banerjee, and Guido Montufar.
Öktem, and Carola-Bibiane Schönlieb. 2024. “Riemannian Geom- 2023. “FoSR: First-Order Spectral Rewiring for Addressing Over-
etry for Efficient Analysis of Protein Dynamics Data.” Proceedings squashing in GNNs.” In The Eleventh International Conference on
of the National Academy of Sciences 121(33): e2318951121. Learning Representations.
Prakash Dwivedi, Vijay, Anh Tuan Luu, Thomas Laurent, Yoshua Kashinath, K., M. Mustafa, A. Albert, J. L. Wu, C. Jiang, S.
Bengio, and Xavier Bresson. 2022. “Graph Neural Networks Esmaeilzadeh, K. Azizzadenesheli, R. Wang, A. Chattopadhyay,
With Learnable Structural and Positional Representations.” In A. Singh, et al. 2021. “Physics-Informed Machine Learning: Case
International Conference on Learning Representations. Studies for Weather and Climate Modelling.” Philosophical trans-
Dym, Nadav and Haggai Maron. 2020. “On the Universality of actions. Series A, Mathematical, Physical, and Engineering Sciences
Rotation Equivariant Point Cloud Networks.” arXiv preprint 379(2194): 20200093–20200093.
arXiv:2010.02449 Kearns, Michael. 1998. “Efficient Noise-Tolerant Learning From
Federer, Herbert. 1959. “Curvature Measures.” Transactions of the Statistical Queries.” Journal of the ACM (JACM) 45(6): 983–1006.
American Mathematical Society 93(3): 418–91. Kiani, Bobak, Lukas Fesser, and Melanie Weber. 2024a. Unitary Con-
Feng, Amy and Melanie Weber. 2024. “Graph Pooling Via Ricci volutions for Learning on Graphs and Groups. Advances in Neural
Flow.” Transactions on Machine Learning Research Information Processing Systems 37 (NeurIPS)
Fesser, Lukas, and Melanie Weber. 2023. “Mitigating Over- Kiani, Bobak T., Thien Le, Hannah Lawrence, Stefanie Jegelka, and
Smoothing and Over-Squashing Using Augmentations of Melanie Weber. 2024b. “On the Hardness of Learning Under Sym-
Forman-Ricci Curvature.” In Learning on Graphs Conference. metries.” In International Conference on Learning Representations.
Fesser, Lukas, and Melanie Weber. 2024. “Effective Structural Encod- Kiani, Bobak T., Jason Wang, and Melanie Weber. 2024c. “Hardness
ings Via Local Curvature Profiles.” In International Conference on of Learning Neural Networks Under the Manifold Hypothesis.”
Learning Representations. Advances in Neural Information Processing Systems 37 (NeurIPS).
Fesser, Lukas, Sergio Serrano de Haro Iváñez, Karel Devriendt, Kipf, Thomas N. and Max Welling. 2017. “Semi-Supervised Classifi-
Melanie Weber, and Renaud Lambiotte. 2023. “Augmentations of cation With Graph Convolutional Networks.” In Proceedings of the
Forman’s Ricci Curvature and Their Applications in Community 5th International Conference on Learning Representations, ICLR
Detection.” Journal of Physics: Complexity 5(3) ’17. ICLR.
Fiedler, Miroslav. 1973. “Algebraic Connectivity of Graphs.” Klimovskaia, Anna, David Lopez-Paz, Léon Bottou, and Maximilian
Czechoslovak Mathematical Journal 23(2): 298–305. Nickel. June 2020. “Poincaré Maps for Analyzing Complex Hier-
Finzi, Marc, Max Welling, and Andrew Gordon Gordon Wilson. archies in Single-Cell Data.” Nature Communications 11(1): 2966.
2021. “A Practical Method for Constructing Equivariant Multi- ISSN 2041-1723.
layer Perceptrons for Arbitrary Matrix Groups.” In Proceedings Kondor, Risi, and Shubhendu Trivedi. 2018. “On the Generaliza-
of the 38th International Conference on Machine Learning, 3318- tion of Equivariance and Convolution in Neural Networks to
28. PMLR. the Action of Compact Groups.” In International Conference on
Forman, Robin. 2003. “Bochner’s Method for Cell Complexes and Machine Learning, 2747–55. PMLR.
Combinatorial Ricci Curvature.” Discrete & Computational Geom- Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. “Ima-
etry 29(3): 323–74. genet Classification With Deep Convolutional Neural Networks.”
Ganea, Octavian, Lagnajit Pattanaik, Connor Coley, Regina Barzilay, In Advances in Neural Information Processing Systems, Vol. 25. MIT
Klavs Jensen, William Green, and Tommi Jaakkola. 2021. “Geo- Press.
mol: Torsional Geometric Generation of Molecular 3d Conformer Kuhn, H. 1955. “The Hungarian Method for the Assignment Prob-
Ensembles.” Advances in Neural Information Processing Systems lem.” Naval Research Logistics Quarterly 2(1-2): 83–97.
34: 13757–69. Lecun, Y., L. Bottou, Y. Bengio, and P. Haffner. 1998. “Gradient-Based
Gori, Monfardini, and Scarselli. “A New Model for Learning in Graph Learning Applied to Document Recognition.” Proceedings of the
Domains.” In Proceedings. 2005 IEEE international joint conference IEEE 86(11): 2278–324.
on neural networks, Vol 2. 729–34. IEEE. Li, Kovachki, Azizzadenesheli, Liu, Bhattacharya, Stuart, and
Gosztolai, Adam, and Alexis Arnaudon. 2021. “Unfolding the Mul- Anandkumar. 2021. “Fourier Neural Operator for Parametric
tiscale Structure of Networks With Dynamical Ollivier-Ricci Partial Differential Equations.” In International Conference on
Curvature.” Nature Communications 12(1): 4561. Learning Representations.
Grant, Michael, Stephen Boyd, and Yinyu Ye. 2006. Disciplined Li, Qimai, Zhichao Han, and Xiao-Ming Wu. “Deeper Insights Into
Convex Programming. Springer. Graph Convolutional Networks for Semi-Supervised Learning.” In
Hamilton, Will, Zhitao Ying, and Jure Leskovec. 2017. “Inductive Proceedings of the AAAI Conference on Artificial Intelligence, Vol.
Representation Learning on Large Graphs.” Advances in neural 32. 3538–45. AAAI Press.
information processing systems 30: 1025–35. Martínez-Rubio, David, Christophe Roux, Christopher Criscitiello,
Hein, Matthias, Jean-Yves Audibert, and Ulrike von Luxburg. 2005. and Sebastian Pokutta. 2023. “Accelerated Methods for Rie-
“From Graphs to Manifolds—Weak and Strong Pointwise Con- mannian Min-Max Optimization Ensuring Bounded Geometric
sistency of Graph Laplacians.” In Learning Theory, Lecture Notes Penalties.” arXiv preprint arXiv:2305.16186.
in Computer Science, edited by Peter Auer and Ron Meir, 470–85, Mei, Song, Theodor Misiakiewicz, and Andrea Montanari. 15–19 Aug
Berlin, Heidelberg, Springer. 2021. “Learning With Invariances in Random Features and Kernel
10 of 11 AI MAGAZINE

Models.” In Proceedings of Thirty Fourth Conference on Learning van der Hoorn, Pim, William J. Cunningham, Gabor Lippner, Carlo
Theory, edited by Mikhail Belkin and Samory Kpotufe, Proceedings Trugenberger, and Dmitri Krioukov. 2021. “Ollivier-Ricci Curva-
of Machine Learning Research, Vol. 134. 3351–418. PMLR. ture Convergence in Random Geometric Graphs.” Physical Review
Morris, Christopher, Martin Ritzert, Matthias Fey, William L. Research 3(1): 013211.
Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Veličković, Petar, Guillem Cucurull, Arantxa Casanova, Adriana
2019. “Weisfeiler and Leman Go Neural: Higher-Order Graph Romero, Pietro Lio, and Yoshua Bengio. 2017. “Graph Attention
Neural Networks.” In Proceedings of the AAAI Conference on Networks.” arXiv preprint arXiv:1710.10903.
Artificial Intelligence. Vol. 33, 4602–9. AAAI Press. Villar, Soledad, David W. Hogg, Kate Storey-Fisher, Weichi Yao,
Nguyen, Khang, Nong Minh Hieu, Vinh Duc Nguyen, Nhat Ho, and Ben Blum-Smith. 2021. “Scalars are Universal: Equivariant
Stanley Osher, and Tan Minh Nguyen. July 23–29 2023. “Revis- Machine Learning, Structured Like Classical Physics.” Advances
iting Over-Smoothing and Over-Squashing Using Ollivier-Ricci in Neural Information Processing Systems 34: 28848–63.
Curvature.” In Proceedings of the 40th International Conference on Von Luxburg, Ulrike. 2007. “A Tutorial on Spectral Clustering.”
Machine Learning, Proceedings of Machine Learning Research, Vol. Statistics and Computing 17: 395–416.
202, 25956–79. PMLR. Wang, Dian, Robin Walters, Xupeng Zhu, and Robert Platt. 2022.
Ni, Chien-Chun, Yu-Yao Lin, Feng Luo, and Jie Gao. 2019. “Commu- “Equivariant 𝑞 Learning in Spatial Action Spaces.” In Conference
nity Detection on Networks With Ricci Flow.” Scientific Reports on Robot Learning, 1713–23. PMLR.
9(1): 9984. Wang, Rui, Robin Walters, and Rose Yu. 2021. “Incorporating Sym-
Ollivier, Yann. 2010. “A Survey of Ricci Curvature for Metric Spaces metry Into Deep Dynamics Models for Improved Generalization.”
and Markov Chains.” In Probabilistic Approach to Geometry, 343– In International Conference on Learning Representations.
81. Mathematical Society of Japan. Watson, Joseph L., David Juergens, Nathaniel R. Bennett, Brian L.
Sia, Jayson, Edmond Jonckheere, and Paul Bogdan. 2019. “Ollivier- Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J.
Ricci Curvature-Based Method to Community Detection in Com- Borst, Robert J. Ragotte, Lukas F. Milles, et al. 2023. “De Novo
plex Networks.” Scientific Reports 9(1): 9800. Design of Protein Structure and Function With RFdiffusion.”
Sinkhorn, Richard, and Paul Knopp. 1967. “Concerning Nonnega- Nature 620(7976): 1089–100.
tive Matrices and Doubly Stochastic Matrices.” Pacific Journal of Weber, Melanie. 2024. “Exploiting Data Geometry in Machine Learn-
Mathematics 21(2): 343–48. ing.” Proceedings of the AAAI Conference on Artificial Intelligence
Spielman, Daniel A., and Shang-Hua Teng. 1996. “Spectral Par- 38(20): 22681–22681.
titioning Works: Planar Graphs and Finite Element Meshes.” Weber, Melanie, and Suvrit Sra. 2021. “Projection-Free Nonconvex
In Proceedings of 37th Conference on Foundations of Computer Stochastic Optimization on Riemannian Manifolds.” IMA Journal
Science, 96–105. IEEE. of Numerical Analysis 42(4): 3241–71.
Subramanian, Akshay, Wenhao Gao, Regina Barzilay, Jeffrey C. Weber, Melanie and Suvrit Sra. 2022. “Riemannian Optimization
Grossman, Tommi Jaakkola, Stefanie Jegelka, Mingda Li, Ju Li, via Frank–Wolfe Methods.” Mathematical Programming 199: 525–
Wojciech Matusik, Elsa Olivetti, Connor W. Coley, et al. March 56.
27 2024. “Closing the Execution Gap in Generative AI for Chemi- Weber, Melanie and Suvrit Sra. 2023. “Global Optimality for
cals and Materials: Freeways or Safeguards.” An MIT Exploration Euclidean CCCP Under Riemannian Convexity.” In International
of Generative AI, [Link] Conference on Machine Learning.
Tahmasebi, Behrooz, and Stefanie Jegelka. 2023. “The Exact Sam- Weber, Melanie, Emil Saucan, and Jürgen Jost. 2017a. “Charac-
ple Complexity Gain From Invariances for Kernel Regression.” terizing Complex Networks With Forman-Ricci Curvature and
Advances in Neural Information Processing Systems 36: 55616–46. Associated Geometric Flows.” Journal of Complex Networks 5(4):
Tian, Yu, Zachary Lubberts, and Melanie Weber. 03 Dec 2023a. 527–50.
“Mixed-Membership Community Detection Via Line Graph Cur- Weber, Melanie, Emil Saucan, and Jürgen Jost. “Coarse Geometry of
vature.” In Proceedings of the 1st NeurIPS Workshop on Symmetry Evolving Networks.” Journal of Complex Networks 6(5): 706–32.
and Geometry in Neural Representations, Proceedings of Machine Weber, Melanie, Johannes Stelzer, Emil Saucan, Alexander
Learning Research, Vol. 197, 219–33. PMLR. Naitsat, Gabriele Lohmann, and Jürgen Jost. 2017c. “Curvature-
Tian, Yu, Zachary Lubberts, and Melanie Weber. 2023b. “Curvature- Based Methods for Brain Network Analysis.” Technical Report
Based Clustering on Graphs.” arXiv:2307.10155. (arXiv:1707.00180).
Topping, Jake, Francesco Di Giovanni, Benjamin Paul Chamberlain, Weber, Melanie, Manzil Zaheer, Ankit Singh Rawat, Aditya Menon,
Xiaowen Dong, and Michael M. Bronstein. 2022. “Understanding and Sanjiv Kumar. 2020. “Robust Large-Margin Learning in
Over-Squashing and Bottlenecks on Graphs Via Curvature.” In Hyperbolic Space.” Advances in Neural Information Processing
International Conference on Learning Representations. Systems 33 (NeurIPS). Cambridge, MA, USA: MIT Press.
Townsend, James, Niklas Koep, and Sebastian Weichwald. 2016. Xu, Keyulu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018.
“Pymanopt: A Python Toolbox for Optimization on Manifolds “How Powerful are Graph Neural Networks?” arXiv preprint
Using Automatic Differentiation.” Journal of Machine Learning arXiv:1810.00826.
Research 17(137): 1–5. Zaheer, Manzil, Satwik Kottur, Siamak Ravanbakhsh, Barnabas
Trillos, Nicolas Garcia and Melanie Weber. “Continuum Limits of Poczos, Russ R. Salakhutdinov, and Alexander J. Smola. 2017.
Ollivier’s Ricci Curvature on Data Clouds: Pointwise Consistency “Deep Sets.” Advances in Neural Information Processing Systems
and Global Lower Bounds.” arXiv:2307.02378. 30: 3394–404.
Udriste, Constantin. 1994. Convex Functions and Optimization Meth- Zhang, Hongyi, and Suvrit Sra. 2016. “First-Order Methods for
ods on Riemannian Manifolds, Vol. 297, Springer Science & Geodesically Convex Optimization.” Conference on Learning The-
Business Media. ory 49: 1617–38.
AI MAGAZINE 11 of 11

Zhang, Hongyi, Sashank J. Reddi, and Suvrit Sra. 2016. “Riemannian AU T H O R B I O G R A P H Y

SVRG: Fast Stochastic Optimization On Riemannian Manifolds.”
Advances in Neural Information Processing Systems 29: 2893–901.
Melanie Weber is an Assistant Professor of Applied
Zhao, Lingxiao, Wei Jin, Leman Akoglu, and Neil Shah. “From Stars
Mathematics and of Computer Science at Harvard
to Subgraphs: Uplifting any GNN With Local Structure Aware-
ness.” In International Conference on Learning Representations. University, where she leads the Geometric Machine
ICLR. Learning Group. Her research studies geometric struc-
Zhao, Zhizhen and Amit Singer. 2014. “Rotationally Invariant Image ture in data and how to leverage such information for
Representation for Viewing Direction Classification in Cryo-EM.” the design of new, efficient Machine Learning algo-
Journal of Structural Biology 186(1): 153–66. rithms with provable guarantees. Previously, she was a
Hooke Research Fellow at the Mathematical Institute
in Oxford and received her PhD from Princeton Uni-
How to cite this article: Weber, M. 2025. versity. She is a recipient of the IMA Leslie Fox Prize in
“Geometric Machine Learning.” AI Magazine 46: Numerical Analysis and a Sloan Research Fellowship
e12210. [Link] in Mathematics.