Abstract:The study of high dimensional data sets often rely on their low dimensional projections that preserve the local geometry of the original space. While numerous methods have been developed to summarize this space as variations of tree-like structures, they are usually non-parametric and "static" in nature. As data may come from systems that are dynamical such as a differentiating cell, a static, non-parametric characterization of the space may not be the most appropriate. Here, we developed a framework, the principal flow, that is capable of characterizing the space in a dynamical manner. The principal flow, defined using neural ODEs, directs motion of a particle through the space, where the trajectory of the particle resembles the principal curve of the dataset. We illustrate that our framework can be used to characterize shapes of various complexities, and is flexible to incorporate summaries of relaxation dynamics.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to construct an interpretable principal curve in high - dimensional datasets that can dynamically represent the data manifold. Traditional methods are usually non - parametric and static and cannot capture the underlying dynamic characteristics in the data. Specifically: 1. **Limitations of Static Non - parametric Methods**: Existing low - dimensional projection methods are usually static and non - parametric and cannot well describe data from dynamic systems, such as changes in the cell differentiation process. 2. **Lack of Dynamic Characteristics**: For data from dynamic systems (such as cells at different differentiation stages), static methods may not be sufficient to capture their time - evolution characteristics. 3. **Lack of Flexibility**: Existing methods are difficult to flexibly handle data distributions of complex shapes and are difficult to incorporate additional information such as relaxation dynamics. To solve these problems, the author has developed a framework based on Neural ODEs - **principal flow**, which can represent the data space in a dynamic way and generate trajectories similar to the principal curve. Through this method, state transitions in complex biological systems can be better understood and predicted. ### Specific Objectives - **Construct Dynamic Representation**: Dynamically represent the spatial structure of high - dimensional data through the principal flow defined by Neural ODEs. - **Handle Complex Geometric Shapes**: Demonstrate that this framework can be used to represent data distributions of various complex shapes and is flexible. - **Combine Dynamic Information**: Incorporate information such as relaxation dynamics into the model to enhance the understanding of the system. - **Predict Perturbation Effects**: Use the Finite - Time Lyapunov Exponent (FTLE) to analyze the impact of perturbations on the system, thereby identifying the areas most sensitive to perturbations. ### Method Overview - **Neural ODEs**: Use a neural network to model the velocity field \(\frac{d\vec{x}}{dt}=g(\vec{x})\), where \(g\) is a neural network, the input is the position of the particle, and the output is its direction of motion. - **Constraints**: To simplify the calculation, the velocity in the velocity field is set to be constant, that is, \(\|\frac{d\vec{x}}{dt}\| = 1\). - **Loss Function**: Train the model by minimizing the distance between the simulated trajectory and the observed data. - **FTLE Analysis**: Calculate the Finite - Time Lyapunov Exponent to evaluate the sensitivity of the system to small perturbations. Through these methods, the author shows how to construct the principal flow on data distributions of different shapes and verifies the effectiveness and flexibility of this method.

Constructing interpretable principal curve using Neural ODEs

Neural Flows: Efficient Alternative to Neural ODEs

Neural Ordinary Differential Equations on Manifolds

Embedding Capabilities of Neural ODEs

Anamnesic Neural Differential Equations with Orthogonal Polynomial Projections

ODE-Inspired Analysis for the Biological Version of Oja's Rule in Solving Streaming PCA

Data-driven optimal control with neural network modeling of gradient flows

Neural modal ordinary differential equations: Integrating physics-based modeling with neural ordinary differential equations for modeling high-dimensional monitored structures

Bayesian Neural Ordinary Differential Equations

Learning on Manifolds: Universal Approximations Properties using Geometric Controllability Conditions for Neural ODEs

Neural Operator with Regularity Structure for Modeling Dynamics Driven by SPDEs

Neural Ordinary Differential Equations for Data-Driven Reduced Order Modeling of Environmental Hydrodynamics

Interpretable Spatio-Temporal Embedding for Brain Structural-Effective Network with Ordinary Differential Equation

Constrained or Unconstrained? Neural-Network-Based Equation Discovery from Data

Neural Conjugate Flows: Physics-informed architectures with flow structure

Modeling Trajectories with Neural Ordinary Differential Equations

OnsagerNet: Learning Stable and Interpretable Dynamics Using a Generalized Onsager Principle

Dynamic Principal Subspaces in High Dimensions

From principal curves to granular principal curves

Graph Fourier Neural ODEs: Bridging Spatial and Temporal Multiscales in Molecular Dynamics

Principal Component Flow Map Learning of PDEs from Incomplete, Limited, and Noisy Data