Neural networks are famously biased toward learning smooth, continuous functions. While this "spectral bias" is a feature for many tasks, it becomes a bug when modeling the real world's sharp edges—physical interfaces in fluid dynamics, the sudden opening of a mouth in a 3D scan, or the literal cut of a pair of scissors through fabric.
To bridge this gap, a recurring strategy has emerged across computer vision and physics-informed machine learning: lifting. By embedding low-dimensional, discontinuous problems into higher-dimensional "ambient" or "hyper" spaces, researchers transform abrupt jumps into smooth manifolds that neural networks can easily digest.
The Continuity Paradox
Standard coordinate-based networks, like Physics-Informed Neural Networks (PINNs) or Neural Radiance Fields (NeRFs), struggle with "accuracy walls" when faced with sharp peaks or interfaces. As Jingyong Ying and colleagues note in Accurate adaptive deep learning method for solving elliptic problems, standard architectures fail to maintain precision at the boundaries of discontinuous physical properties.
The intuition behind lifting is to treat a discontinuity not as a break in the function, but as a cross-section of a higher-dimensional continuous shape. Imagine a 1D jump from 0 to 1. In 1D, this is a cliff. But if you view it as a 2D curve winding through a third dimension, the "jump" is simply where the curve passes behind itself from your perspective.
Resolving Topological Ambiguity
In dynamic 3D reconstruction, the most difficult moments occur when surfaces touch or separate—a phenomenon called topological change. When a person opens their mouth, the lips "tear" apart. A standard 3D deformation field struggles here because it tries to stretch a single continuous surface until it snaps, often resulting in "ghosting" artifacts.
To solve this, Simon Giebenhain and Matthias Nießner in MonoNPHM augment deformation fields with additional "hyper-dimensions." These extra coordinates allow the model to resolve ambiguities: in the 3D world, the upper and lower lips are touching, but in the higher-dimensional ambient space, they are distinct and distant. This prevents the collapse of separate geometric features into the same coordinate space.
Similarly, Chengwei Zheng and Feng Xu use this mechanism in EditableNeRF to handle objects separating or joining. By mapping key points into a higher-dimensional space, complex structural breaks are represented as continuous paths, maintaining structural integrity during user edits.
Lifting for Physics and Simulation
This concept extends beyond visual reconstruction into hard physics. In XI-DeepONet, Ran Bi and Weibing Deng solve elliptic interface problems by incorporating level set functions as an additional coordinate. This transforms a discontinuous $d$-dimensional problem into a continuous $(d+1)$-dimensional space, allowing the network to approximate sharp transitions without expensive mesh generation.
Perhaps the most intuitive application is found in Wind Lifter. Mengfei Liu and Eitan Grinspun simulate the cutting of fabric by lifting 2D coordinates into 3D using "winding numbers." A cut that would be a sharp discontinuity in 2D becomes a smooth, continuous volumetric field in 3D. When the simulation needs to know the state of the fabric, it simply "restricts" the 3D field back to the 2D domain, where the jump naturally reappears.
Why It Works
The success of high-dimensional lifting relies on two factors:
- Neural Bias Alignment: It aligns the problem with the neural network's natural tendency toward smoothness.
- Manifold Unfolding: It provides the mathematical "elbow room" needed to represent complex motions—like objects appearing, disappearing, or folding—that violate standard 3D manifold assumptions.
By moving from 3D to "Hyper-3D," models like DyLiN and DynamicSurf can represent non-rigid deformations and topological shifts that are mathematically impossible to capture in lower-dimensional, purely continuous mappings.