About my research

Neuromanifolds: the geometry of model expressivity

Given a parametrization Ψ: Θ → 𝓕, its image is a neuromanifold.

The geometry induced by this parametrization shapes optimization dynamics.

  • Neural networks
  • Nonnegative tensor factorization
  • The principal-minor map

I am working on questions at the intersection of algebraic geometry and machine learning. More concretely, I am interested in how to apply algebro-geometric tools to understand the geometry of function spaces of algebraic neural networks.

For a fixed neural network architecture, the network parameters map into an ambient space, and the image is a semi-algebraic set called a neuromanifold. The geometry and topology of the neuromanifold affect the training dynamics of neural networks. I am interested in how topology, geometry, and invariants can improve our understanding of black-box neural network training and potentially bring practical advantages.

Publications

  1. Submitted

    Algebraic geometry of rational neural networks

    Abstract

    A geometric study of shallow rational neural networks and the function spaces they represent, with singularities providing structure that can be investigated through algebraic geometry.

    Editorial summary
  2. Submitted

    Expressivity of Shallow Neural Networks Over Finite Fields

    Abstract

    An investigation of the expressivity and geometry of shallow neural networks over finite fields—a discrete setting that also offers a model for networks with low-precision weights.

    Editorial summary
  3. Published

    Sign patterns of principal minors of real symmetric matrices

    Abstract

    A study of the sign patterns that can occur among the principal minors of real symmetric matrices, viewed through the structure of a constrained real-algebraic problem.

    Editorial summary
  4. Published

    Geometry of Rank Constraints in Shallow Polynomial Neural Networks

    Abstract

    A connection between shallow polynomial neural networks, tensor decomposition, and the geometry of rank constraints arising from monomial activations.

    Editorial summary
  5. Published

    Likelihood Geometry of Determinantal Point Processes

    Abstract

    An examination of the likelihood geometry of determinantal point processes through the algebraic structure of their statistical models.

    Editorial summary
  6. Published

    Chromatic Graph Homology for Brace Algebras

    Abstract

    A development of chromatic graph homology in the setting of brace algebras, bringing graph-theoretic constructions into conversation with algebraic operations.

    Editorial summary

The neural network framework

  1. Step 1

    Fix the architecture

    Depth, width, and connectivity stay fixed.

  2. Step 2

    Choose an activation function

    Polynomial, rational, ReLU, or tropical.

  3. Step 3

    Define the parameter map

    w fw

Image in function space

Neuromanifold

equations · fibers · topology

  1. I

    Polynomial networks

    Monomial activations connect neural-network function spaces with tensor decomposition and simultaneous Waring rank.

  2. II

    Rational networks

    Rational activations lead toward Chow varieties and toward questions about learning the location and type of singularities from data.

  3. III

    Finite fields

    Finite-field networks offer a discrete setting for studying expressivity, geometry, and low-precision weights.

  4. IV

    ReLU neural networks

    Prony varieties and piecewise-linear geometry.

  5. V

    Tropical neural networks

    Tropical geometry and out-of-distribution generalization.