Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

Neural manifold hypothesis suggests activity of neural population forms a low-dimensional manifold
Dimensionality reduction techniques don’t provide explicit parameterization of manifold or capture global structure
Topological data analysis methods reveal shared topological structure between neural manifolds and task variables
Leverage tools from Riemannian geometry and topologically-aware deep generative models to study geometry of neural manifolds
Computes explicit parameterization and estimates local extrinsic curvature

Machine learning uses the manifold hypothesis to explain real-world data
Neural population activity is hypothesized to form low-dimensional manifolds
Dimensionality reduction techniques can reveal lower-dimensional structure
Topological data analysis can reveal shared topological structure
Few methods exist to explicitly quantify and parameterize the geometric structure of neural manifolds
This paper introduces a novel approach to study the geometry of neural manifolds

Topological methods are robust to continuous deformations of the neural manifold.
Interested in quantifying and modeling nonlinear deformations of a population of neurons.
Hypothesize that curvature imposed by deformations may correlate with perceptual or behavioral variables.

VAEs are a type of generative deep latent variable model used for unsupervised learning
VAEs encode input data in a compact latent representation and then decode the latent representation to reconstruct the original input data
The latent variables take value in Euclidean space, z ∈ R L where L < N
The neural manifold of interest is represented by an immersion f going from the template manifold M * to the neural manifold M ⊂ R N
The map f parameterizes the manifold f (S 1 ) immersed in R N with angular coordinates in S 1
The immersion f induces a Riemannian metric structure on M through the first fundamental form
The Riemannian metric of the neural manifold allows us to calculate geometric quantities like angles, distances, areas and various types of curvatures
Riemannian geometry of deep learning has focused on intrinsic notions of curvatures
We suggest instead to use an extrinsic notion of curvature, such as the mean curvature vector
The norms of the mean curvatures of the circle, torus and sphere are given in Example 1

Propose a method to compute pullback metric, second fundamental form, and mean curvature vector of neural manifolds
Method uses a variational autoencoder
Topology of neural manifold determined with existing topological data analysis methods
Template manifold chosen to be simplest smooth manifold homeomorphic to neural manifold
Mapping f from template manifold to neural manifold learned with tailored variational autoencoder
VAE is “topologically-aware” with latent space constrained to be template manifold
VAE trained with tailored loss function
Mean curvature H of M computed with map f
VAE pipeline for hyperspheres and 2-torus implemented
Latent space constrained to be template manifold
Kullback-Leibler divergence with posterior and prior distributions defined on template manifold
Geodesic loss between latent variables and task variables
Automatic differentiation used to compute pullback metric and curvatures from f
VAE implementation uses Geomstats software

Synthetic datasets of circles, spheres, and tori are distorted by small Gaussian “bumps” and embedded in a high-dimensional space
Random N-dimensional rotation and i.d.d. Gaussian noise are added
Parameter α modulates the amplitude around the ring, creating extrinsic curvature
Parameter α introduces curvature in the vicinity of the north and south poles of S 2
Parameter α introduces extrinsic curvature by stretching the torus on opposite sides
Mean curvature vector is calculated from true synthetic immersion and the learned immersion
Normalized error is calculated and varies between 0.5% and 5% of the actual curvature for noise levels that match real neural data

Our approach estimates the geometry of synthetic neural manifolds generated from smooth deformations of circles, spheres, and tori.
These manifolds have topology that is highly relevant across multiple sensory modalities.
Lemma 5 states that two immersions of a smooth manifold can be related by a rotation and translation.
The pullback metric of a circle is a 1x1 matrix.
The pullback metric of a sphere is a 2x2 matrix.
The pullback metric of a torus is a 2x2 matrix.
Intrinsic curvature makes no reference to a higher dimensional embedding space.
The Riemann curvature tensor, Ricci curvature tensor, and scalar curvature of a one-dimensional manifold are all 0.
The scalar curvatures of the neural manifolds from Fig. 1 are computed in (Cox, 2019).
The second fundamental form of a surface in R3 is a matrix of second partial derivatives of a twice differentiable function.
The mean curvature vector of a surface in R3 is defined by its second fundamental form.
The mean curvature vector of a surface in R3 can also be defined by the curvature of a plane curve.
The mean curvature vector of a submanifold of N is defined by the trace of the inverse of the Riemannian metric matrix.
The mean curvature vector of an immersed manifold is defined by the trace of the inverse of the Riemannian metric matrix.
The mean curvature vector of a circle immersed in R N is given.