118 Multi-Origin High-Dimensional Geometric Neural Networks

Bosley Zhang
Join to follow...
Follow/Unfollow Writer: Bosley Zhang
By following, you’ll receive notifications when this author publishes new articles.
Don't wait! Sign up to follow this writer.
WriterShelf is a privacy-oriented writing platform. Unleash the power of your voice. It's free!
Sign up. Join WriterShelf now! Already a member. Login to WriterShelf.
605   0  
·
2026/04/25
·
4 mins read
☕


Reconstruction of Neural Network Architectures under the Multi-Origin High-Dimensional Geometry Framework

 

Abstract: Existing neural network architectures built upon fully connected or self-attention mechanisms are mostly grounded in dense matrix operations over flat spaces, suffering from bloated parameter scales and excessive computational power consumption. Drawing on the framework of Multi-Origin High-Dimensional Geometry (MOC), this paper proposes a theoretical reconstruction scheme for neural network architectures. Within this scheme, neurons are defined as local origins, while network weights and activation mechanisms are interpreted as curvature coupling and threshold triggering. It seeks to replace global tiled matrix operations with high-dimensional geodesic signal propagation. This work explores the theoretical potential of such geometric structures for cutting computational redundancy and optimising information transmission paths, and offers a geometric perspective for next-generation low-energy computing architectures.

 

I. Introduction: Computational Redundancy in Conventional Architectures

 

Dominant deep learning architectures today, including multilayer perceptrons and Transformers, rely primarily on matrix multiplication and tensor operations. Viewed geometrically, such computations are constructed on a globally unified single-origin coordinate system.

 

To fit complex data distributions, models often require extremely deep stacks and massive parameter counts, resulting in abundant redundant dense connections and repeated iterations between layers. This tiled computational structure inevitably brings surging data movement and bottlenecks in computing energy consumption. Exploring an alternative network paradigm built on a superior geometric structure carries substantial theoretical value.

 

II. MOC Geometric Redefinition of Network Components

 

Within the Multi-Origin High-Dimensional Geometry (MOC) framework, a computational network is no longer treated merely as a two-dimensional matrix array, but as a multi-origin, variable-curvature high-dimensional manifold. Information transfer arises naturally from the evolution of curvature potential differences between origin clusters. Accordingly, ten core components of neural networks may be reinterpreted as follows:

 

1. Neuron: an independent origin in its respective dimension, carrying local features and serving as a dynamic curvature reference.

2. Weight: the high-dimensional association strength between origins, namely the curvature coupling coefficient across origins.

3. Bias: an intrinsic curvature offset belonging to a single origin.

4. Activation function: a threshold-triggering mechanism for origin curvature, which governs the gating of dimensional projections.

5. Forward and backpropagation: forward propagation denotes directed information transmission along high-dimensional geodesics (geometric shortest paths); backpropagation denotes error backtracking and correction along curvature gradients.

6. Loss function: a global aggregate measure of geometric projection deviation across the full set of origin clusters.

7. Gradient descent: an evolutionary process that dynamically adjusts coupling relations between origins by following curvature gradients.

8. Network layer: a collection of origin clusters occupying distinct hierarchical levels in the high-dimensional space.

9. Feature mapping: the projective representation of high-dimensional geometric structures within a low-dimensional observation space.

10. Attention mechanism: a process that dynamically selects high-contribution origins and reinforces critical curvature associations.

 

III. Theoretical Deduction of Computational Efficiency and Energy Consumption

 

Under the multi-origin geometric framework, network information flow no longer follows dense full-connection matrix multiplication, but geodesic propagation. From the perspectives of computational complexity and information theory, this geometric reconstruction holds the following potential efficiency advantages:

 

1. Reduced computational complexity (sparsification and shortest-path routing)

Traditional matrix operations compute pairwise associations for all nodes, typically with O(N^2) complexity. In the MOC framework, information propagates along high-dimensional geodesics, meaning only origins lying on key curvature gradient paths enter strong coupling. This geometric structure is inherently highly sparse, and may theoretically drastically reduce operation counts and floating-point operations (FLOPs).

2. Lower data movement overhead

Conventional tiled matrix operations impose heavy demands on memory bandwidth. The MOC architecture defines weights as curvature coupling coefficients and attention as dynamic origin selection. Computation is dominated by local updates of topological relations between origins, theoretically cutting data movement between storage and compute units.

3. Energy efficiency mapping

From fundamental principles of computing, fewer operations and less data movement directly reduce energy draw. With a better underlying geometric structure, sparser parameters and more direct signal paths yield higher overall system energy efficiency.

 

IV. Conclusion and Future Work

 

This paper establishes a system of geometric redefinitions for neural network components based on MOC theory. The framework illustrates that when fundamental operations are grounded in multi-origin and high-dimensional curvature, information may propagate along shorter geometric paths, theoretically delivering benefits in reduced computational redundancy and energy use.

 

Limitations and scope: this is a purely theoretical investigation. Geometric optimality does not automatically translate to practical engineering performance, which remains subject to physical constraints such as floating-point precision, hardware organisation and parallel compute efficiency. Future work will focus on translating this geometric framework into concrete discrete operators or tensor-network algorithms, and testing convergence speed, parameter efficiency and energy behaviour on standard benchmark datasets to validate the engineering viability of the proposed architecture.

 



WriterShelf™ is a unique multiple pen name blogging and forum platform. Protect relationships and your privacy. Take your writing in new directions. ** Join WriterShelf**
WriterShelf™ is an open writing platform. The views, information and opinions in this article are those of the author.


Article info

This article is part of:
分類於:
日期:
創作於:2026/04/25,最後更新於:2026/11/13。
合計:767字


Share this article:
About the Author

I love science as much as art, logic as deeply as emotion.

I write the softest human stories beneath the hardest sci-fi.

May words bridge us to kindred spirits across the world.




Join the discussion now!
Don't wait! Sign up to join the discussion.
WriterShelf is a privacy-oriented writing platform. Unleash the power of your voice. It's free!
Sign up. Join WriterShelf now! Already a member. Login to WriterShelf.