Revisiting Transformation Invariant Geometric Deep Learning: Are Initial Representations All You Need?

Ziwei Zhang, Xin Wang, Zeyang Zhang, Peng Cui, Wenwu Zhu

Abstract

Geometric deep learning has achieved significant success in handling geometric data like point clouds and graphs. A critical inductive bias for these models is maintaining invariance toward transformations such as translation, rotation, and scaling. While existing approaches often rely on computationally expensive, custom-built layers or data augmentation, this paper explores whether transformation-invariant and distance-preserving initial representations are sufficient. We propose Transformation Invariant Neural Networks (TinvNN), a general framework that utilizes a modified Multi-Dimensional Scaling (MDS) technique to achieve strict transformation invariance. By normalizing eigenvalues to handle scaling, TinvNN can be combined with any standard neural network backbone. Our results on point cloud analysis and combinatorial optimization demonstrate that TinvNN achieves state-of-the-art performance with minimal architectural complexity.

Introduction

In geometric deep learning, the absolute coordinates of data points are often arbitrary. A 3D model of a chair remains a chair regardless of its orientation, position, or size. However, standard neural networks like PointNet or Graph Neural Networks (GNNs) typically treat transformed coordinates as entirely new inputs. This leads to poor generalization when models encounter rotations or scales not present in the training set.

Historically, this has been addressed through data augmentation, manual feature engineering (e.g., using relative distances), or sophisticated equivariant architectures. These methods are often computationally heavy, lose global structural information, or are difficult to adapt to new baseline architectures. TinvNN addresses these issues by shifting the focus from complex layer design to the initial data representation.

Methodology: The TinvNet Framework

TinvNet acts as a universal pre-processing plug-in that transforms raw coordinates into an invariant space before they reach the neural network backbone. The process involves four key steps:

  1. Distance Matrix Calculation: Pairwise Euclidean distances between all points are computed to capture the geometric structure.
  2. Modified MDS: Classical Multi-Dimensional Scaling is applied to convert the distance matrix into a similarity matrix, followed by eigen-decomposition to find a new coordinate basis.
  3. Eigenvalue Normalization: Standard MDS is invariant to rotation and translation but not scaling. TinvNet introduces eigenvalue normalization to preserve relative distance ratios, ensuring invariance to uniform scaling.
  4. Backbone Integration: The resulting invariant representations are fed into standard backbones like PointNet, DGCNN, or Graph Attention Networks (GAT).

Additionally, the framework can be extended to achieve equivariance by using the Kabsch-Umeyama algorithm at the output layer to reverse the initial transformation mapping.

Experimental Results

Point Cloud Analysis

The model was tested on ModelNet40 (classification) and ShapeNet (segmentation). In a challenging "extreme rotation" scenario—where models are trained on z-aligned objects but tested on arbitrary 3D rotations—standard models fail significantly. For instance, a standard DGCNN's accuracy drops to 22.3%. With the TinvNet plug-in, the same DGCNN achieves 89.5% accuracy, matching or exceeding the performance of specialized, mathematically complex invariant networks.

Qualitative comparison of classification results under arbitrary 3D rotations. Standard models like PointNet and DGCNN misclassify rotated objects, while TinvNet-enhanced versions maintain correct predictions.

Combinatorial Optimization

TinvNet was also applied to the Travelling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP). Standard GAT models show extreme sensitivity to coordinate shifts; for TSP-100, a simple translation causes the error gap to jump from 9.4% to over 570%. TinvNet+GAT maintains a stable 9.8% error gap across all similarity transformations, including rotation, scaling, and reflection.

Discussion and Insights

  • Decoupling Invariance: TinvNN proves that invariance can be achieved at the representation level, allowing researchers to use the most powerful standard architectures without needing to redesign layers for equivariance.
  • Efficiency: The MDS pre-processing is highly efficient, taking less than 50ms for a 2,048-point cloud on a standard GPU.
  • Limitations: The current framework is limited to similarity transformations and does not yet handle non-uniform affine transformations (like shearing). Additionally, the sign ambiguity of eigenvectors requires a small amount of sign-flip data augmentation to ensure the network remains robust to different eigen-bases.

In conclusion, TinvNN provides a simple yet mathematically rigorous baseline for transformation-invariant geometric deep learning, suggesting that sophisticated initial representations may indeed be "all you need."

Summary updated

Abstract

Geometric deep learning, i.e., designing neural networks to handle the ubiquitous geometric data such as point clouds and graphs, have achieved great successes in the last decade. One critical inductive bias is that the model can maintain invariance towards various transformations such as translation, rotation, and scaling. The existing graph neural network (GNN) approaches can only maintain permutation-invariance, failing to guarantee invariance with respect to other transformations. Besides GNNs, other works design sophisticated transformation-invariant layers, which are computationally expensive and difficult to be extended. To solve this problem, we revisit why the existing neural networks cannot maintain transformation invariance when handling geometric data. Our findings show that transformation-invariant and distance-preserving initial representations are sufficient to achieve transformation invariance rather than needing sophisticated neural layer designs. Motivated by these findings, we propose Transformation Invariant Neural Networks (TinvNN), a straightforward and general framework for geometric data. Specifically, we realize transformation-invariant and distance-preserving initial point representations by modifying multi-dimensional scaling before feeding the representations into neural networks. We prove that TinvNN can strictly guarantee transformation invariance, being general and flexible enough to be combined with the existing neural networks. Extensive experimental results on point cloud analysis and combinatorial optimization demonstrate the effectiveness and general applicability of our proposed method. Based on the experimental results, we advocate that TinvNN should be considered a new starting point and an essential baseline for further studies of transformation-invariant geometric deep learning.
BibTeX
@article{zhang2025revisiting,
  author = {Ziwei Zhang and Xin Wang and Zeyang Zhang and Peng Cui and Wenwu Zhu},
  title = {{Revisiting Transformation Invariant Geometric Deep Learning: Are Initial Representations All You Need?}},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year = {2025},
  doi = {10.1109/tpami.2025.3626735},
}
Figure