TLDR: LogViG is a new hybrid CNN-GNN model that introduces Logarithmic Scalable Graph Construction (LSGC) to efficiently build graphs in Vision GNNs. LSGC addresses the computational complexity and over-squashing issues of previous methods by scaling connections logarithmically, enabling global links with fewer connections. Combined with a high-resolution shortcut, LogViG achieves superior accuracy, fewer parameters, and lower computational costs compared to existing CNN, ViT, and ViG architectures on image classification and semantic segmentation tasks.
Vision Graph Neural Networks (ViG) are emerging as powerful tools in computer vision, offering a fresh perspective compared to traditional Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Instead of processing images as grids of pixels or sequences of patches, ViGs represent images as networks of interconnected patches, allowing them to identify objects by understanding relationships between different parts of an image, even if they are far apart.
Despite their promise, ViGs have faced a significant hurdle: computational complexity, especially when constructing the underlying graph structure. Common methods like k-nearest neighbor (KNN) can be very expensive for larger, high-resolution images. Another approach, Sparse Vision Graph Attention (SVGA), simplifies the process but can suffer from ‘over-squashing,’ where too much information is compressed into a single vector, leading to a loss of global context as image resolution increases.
Addressing these challenges, researchers from The University of Texas at Austin – Mustafa Munir, Alex Zhang, and Radu Marculescu – have proposed a novel solution: Logarithmic Scalable Graph Construction (LSGC). This new method aims to enhance performance by intelligently limiting the number of long-range links, scaling connections logarithmically rather than linearly. LSGC generates fewer connections for high-resolution images, mitigating over-squashing, and prioritizes connections closer to a patch while still enabling crucial long-range connections across the entire image. This approach significantly reduces computational complexity while maintaining the ability to capture both local and global features.
Building on LSGC, the team introduces LogViG, a novel hybrid CNN-GNN model. LogViG integrates LSGC into its architecture and also incorporates a high-resolution shortcut (HRS). Inspired by the success of multi-scale and high-resolution architectures, the HRS allows for the fusion of features between high-resolution and low-resolution branches, enabling multi-scale feature interaction and preventing performance degradation often seen when downsampling feature maps.
Extensive experiments have shown that LogViG outperforms existing ViG, CNN, and ViT architectures across various metrics. On image classification tasks using ImageNet-1K, LogViG models demonstrate superior accuracy, fewer parameters, and lower computational costs (GMACs). For instance, their smallest model, Ti-LogViG, achieved a 79.9% top-1 accuracy on ImageNet-1K, which is 1.7% higher than Vision GNN, with a substantial reduction in parameters (24.3%) and GMACs (35.3%). Similarly, on semantic segmentation tasks using ADE20K, LogViG also showed superior performance compared to other state-of-the-art models.
Ablation studies further confirmed the benefits of LSGC and the high-resolution shortcut. Replacing SVGA with LSGC alone resulted in a 0.4% accuracy gain without additional parameters, and adding the high-resolution shortcut provided another 0.1% accuracy increase with minimal parameter overhead. The research also explored network depth versus width, finding that a deeper and narrower LogViG architecture generally yielded better performance.
Also Read:
- Bridging Self-Attention and Convolution with Translution for Enhanced Deep Learning
- Unifying Data Understanding with Chimera: A New Approach to Deep Learning Topology
In conclusion, Logarithmic Scalable Graph Construction (LSGC) and the LogViG architecture represent a significant advancement in the design of Vision GNNs. By efficiently balancing long-range links and computational cost, LogViG offers a powerful and efficient alternative for complex vision tasks. You can read the full research paper here.
