We propose HyperSteiner – an efficient heuristic algorithm for computing Steiner minimal trees in the hyperbolic space. HyperSteiner extends the Euclidean Smith-Lee-Liebman algorithm, which is grounded in a divide-and-conquer approach involving the Delaunay triangulation. The central idea is rephrasing Steiner tree problems with three terminals as a system of equations in the Klein-Beltrami model. Motivated by the fact that hyperbolic geometry is well-suited for representing hierarchies, we explore applications to hierarchy discovery in data. Results show that HyperSteiner infers more realistic hierarchies than the Minimum Spanning Tree and is more scalable to large datasets than Neighbor Joining.

Lattice reduction is a combinatorial optimization problem aimed at finding the most orthogonal basis in a given lattice. The Lenstra–Lenstra–Lovász (LLL) algorithm is the best algorithm in the literature for solving this problem. In light of recent research on algorithm discovery, in this work, we would like to answer this question: is it possible to parametrize the algorithm space for lattice reduction problem with neural networks and find an algorithm without supervised data? Our strategy is to use equivariant and invariant parametrizations and train in a self-supervised way. We design a deep neural model outputting factorized unimodular matrices and train it in a self-supervised manner by penalizing non-orthogonal lattice bases. We incorporate the symmetries of lattice reduction into the model by making it invariant to isometries and scaling of the ambient space and equivariant with respect to the hyperocrahedral group permuting and flipping the lattice basis elements. We show that this approach yields an algorithm with comparable complexity and performance to the LLL algorithm on a set of benchmarks. Additionally, motivated by certain applications for wire-less communication, we extend our method to a convolutional architecture which performs joint reduction of spatially-correlated lattices arranged in a grid, thereby amortizing its cost over multiple lattices.

We study convolutional neural networks with monomial activation functions. Specifically, we prove that their parameterization map is regular and is an isomorphism almost everywhere, up to rescaling the filters. By leveraging on tools from algebraic geometry, we explore the geometric properties of the image in function space of this map - typically referred to as neuromanifold. In particular, we compute the dimension and the degree of the neuromanifold, which measure the expressivity of the model, and describe its singularities. Moreover, for a generic large dataset, we derive an explicit formula that quantifies the number of critical points arising in the optimization of a regression loss.

Human olfactory perception is understudied in the whole spectrum of neuroscience, from computational to system neuroscience. In this study, we explore the hierarchy underlying human olfactory perception by embedding perceptual data in the hyperbolic space. Previous research emphasizes the significance of hyperbolic geometry in gaining insights into the neural encoding of natural odorants. This is due to the exponential growth of the hyperbolic space, that makes it appropriate to encode hierarchical data. We employ a contrastive learning approach over the Poincare ball in order to embed olfactory perceptual data in a hyperbolic space. The results indicate the emergence of a hierarchical representation in the hyperbolic space, which could have implications for understanding the structure of the olfactory perceptual space in the brain. Our finding suggests that the human brain may encode olfactory perceptions in a hierarchical manner, where higher odor perceptual certainty correlates with deeper levels in the hierarchical representation.

In this work, we formally prove that, under certain conditions, if a neural network is invariant to a finite group then its weights recover the Fourier transform on that group. This provides a mathematical explanation for the emergence of Fourier features - a ubiquitous phenomenon in both biological and artificial learning systems. The results hold even for non-commutative groups, in which case the Fourier transform encodes all the irreducible unitary group representations. Our findings have consequences for the problem of symmetry discovery. Specifically, we demonstrate that the algebraic structure of an unknown group can be recovered from the weights of a network that is at least approximately invariant within certain bounds. Overall, this work contributes to a foundation for an algebraic learning theory of invariant neural network representations.

Hyperbolic machine learning is an emerging field aimed at representing data with a hierarchical structure. However, there is a lack of tools for evaluation and analysis of the resulting hyperbolic data representations. To this end, we propose Hyperbolic Delaunay Geometric Alignment (HyperDGA) - a similarity score for comparing datasets in a hyperbolic space. The core idea is counting the edges of the hyperbolic Delaunay graph connecting datapoints across the given sets. We provide an empirical investigation on synthetic and real-life biological data and demonstrate that HyperDGA outperforms the hyperbolic version of classical distances between sets. Furthermore, we showcase the potential of HyperDGA for evaluating latent representations inferred by a Hyperbolic Variational Auto-Encoder.

Spaces of data naturally carry intrinsic geometry. Statistics and machine learning can leverage on this rich structure in order to achieve efficiency and semantic generalization. Extracting geometry from data is therefore a fundamental challenge which by itself defines a statistical, computational and unsupervised learning problem. To this end, symmetries and metrics are two fundamental objects which are ubiquitous in continuous and discrete geometry. Both are suitable for data-driven approaches since symmetries arise as interactions and are thus collectable in practice while metrics can be induced locally from the ambient space. In this thesis, we address the question of extracting geometry from data by leveraging on symmetries and metrics. Additionally, we explore methods for statistical inference exploiting the extracted geometric structure. On the metric side, we focus on Voronoi tessellations and Delaunay triangulations, which are classical tools in computational geometry. Based on them, we propose novel non-parametric methods for machine learning and statistics, focusing on theoretical and computational aspects. These methods include an active version of the nearest neighbor regressor as well as two high-dimensional density estimators. All of them possess convergence guarantees due to the adaptiveness of Voronoi cells. On the symmetry side, we focus on representation learning in the context of data acted upon by a group. Specifically, we propose a method for learning equivariant representations which are guaranteed to be isomorphic to the data space, even in the presence of symmetries stabilizing data. We additionally explore applications of such representations in a robotics context, where symmetries correspond to actions performed by an agent. Lastly, we provide a theoretical analysis of invariant neural networks and show how the group-theoretical Fourier transform emerges in their weights. This addresses the problem of symmetry discovery in a self-supervised manner.  

Datamängder innehar en naturlig inneboende geometri. Statistik och maskininlärning kan dra nytta av denna rika struktur för att uppnå effektivitet och semantisk generalisering. Att extrahera geometri ifrån data är därför en grundläggande utmaning som i sig definierar ett statistiskt, beräkningsmässigt och oövervakat inlärningsproblem. För detta ändamål är symmetrier och metriker två grundläggande objekt som är allestädes närvarande i kontinuerlig och diskret geometri. Båda är lämpliga för datadrivna tillvägagångssätt eftersom symmetrier uppstår som interaktioner och är därmed i praktiken samlingsbara medan metriker kan induceras lokalt ifrån det omgivande rummet. I denna avhandling adresserar vi frågan om att extrahera geometri ifrån data genom att utnyttja symmetrier och metriker. Dessutom utforskar vi metoder för statistisk inferens som utnyttjar den extraherade geometriska strukturen. På den metriska sidan fokuserar vi på Voronoi-tessellationer och Delaunay-trianguleringar, som är klassiska verktyg inom beräkningsgeometri. Baserat på dem föreslår vi nya icke-parametriska metoder för maskininlärning och statistik, med fokus på teoretiska och beräkningsmässiga aspekter. Dessa metoder inkluderar en aktiv version av närmaste grann-regressorn samt två högdimensionella täthetsskattare. Alla dessa besitter konvergensgarantier på grund av Voronoi-cellernas anpassningsbarhet. På symmetrisidan fokuserar vi på representationsinlärning i sammanhanget av data som påverkas av en grupp. Specifikt föreslår vi en metod för att lära sig ekvivarianta representationer som garanteras vara isomorfa till datarummet, även i närvaro av symmetrier som stabiliserar data. Vi utforskar även tillämpningar av sådana representationer i ett robotiksammanhang, där symmetrier motsvarar handlingar utförda av en agent. Slutligen tillhandahåller vi en teoretisk analys av invarianta neuronnät och visar hur den gruppteoretiska Fouriertransformen framträder i deras vikter. Detta adresserar problemet med att upptäcka symmetrier på ett självövervakat sätt.

Relative representations are an established approach to zero-shot model stitching, consisting of a non-trainable transformation of the latent space of a deep neural network. Based on insights of topological and geometric nature, we propose two improvements to relative representations. First, we introduce a normalization procedure in the relative transformation, resulting in invariance to non-isotropic rescalings and permutations. The latter coincides with the symmetries in parameter space induced by common activation functions. Second, we propose to deploy topological densification when fine-tuning relative representations, a topological regularization loss encouraging clustering within classes. We provide an empirical investigation on a natural language task, where both the proposed variations yield improved performance on zero-shot model stitching.