HexNets: Exploration of Hexagonal Neural Network Architecture

HexNets is an open-source project exploring the implementation of neural networks with hexagonal grid structures. It challenges the traditional rectangular grid convolution approach and provides a new deep learning architecture option for processing data with hexagonal symmetry (such as satellite images and game maps). This article will introduce it from aspects including background motivation, grid advantages, technical challenges, application scenarios, research significance, and future directions.

In the field of deep learning, Convolutional Neural Networks (CNNs) are standard tools for processing images and spatial data, but mainstream implementations are all based on rectangular grid structures. The HexNets project raises the question: What changes would occur if the underlying grid of neural networks is changed to hexagonal? It aims to explore architectures more suitable for non-rectangular symmetric data.

Hexagonal grids are common in nature and engineering, and have three major advantages over rectangular grids:

1. Isotropy: Six equidistant neighbors, no horizontal or vertical bias;
2. Better sampling efficiency: Fewer sampling points to cover the same area, suitable for spherical data;
3. Better adjacency relations: Six neighbors provide richer local context, helping to capture complex spatial patterns.

Migrating to hexagonal grids requires solving multiple technical challenges:

• Coordinate system design: Adopt axial or cube coordinates (mapping to 3D plane facilitates distance calculation);
• Redefinition of convolution operations: Hexagonal convolution kernels (e.g., radius 1 with 7 units) change the receptive field;
• Pooling and upsampling: Pooling and upsampling operations adapted to hexagonal topology;
• Framework integration: Need to customize CUDA kernels or index transformations to adapt to frameworks like PyTorch/TensorFlow.

Hexagonal neural networks can be applied in:

• Earth observation and satellite images: Avoid distortion from rectangular resampling;
• Strategy game AI: Naturally adapt to situation evaluation of hexagonal maps;
• Computational fluid dynamics and physical simulation: Improve numerical stability;
• Molecular structure modeling: More suitable for hexagonal symmetric structures (such as benzene rings, graphene).

The value of HexNets lies in:

• Challenging the default assumptions of deep learning and exploring basic architecture choices as hyperparameters;
• Similar to GNN breaking grid limits and Transformer breaking sequence limits, showing the possibility of new basic structures;
• Helping to understand the inductive bias of neural networks and design more efficient dedicated architectures.

HexNets can develop in the future in:

• Performance optimization: Develop dedicated CUDA kernels to improve efficiency;
• Hybrid architectures: Combine rectangular and hexagonal convolution layers;
• Spherical expansion: Implement distortion-free global data learning;
• Benchmarking: Establish standardized tests to quantify pros and cons.

HexNets is an imaginative open-source project that explores new possibilities of neural networks from geometric structures. Although it will not replace mainstream rectangular convolution in the short term, it provides new ideas for specific fields, is a valuable starting point for processing hexagonal data, and demonstrates the vitality of the open-source community in architectural innovation.