Point Grid Convolution With Dynamic Kernels for Graph Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph convolutional networks (GCNs) face performance limitations and irregular workloads when processing graph-structured data, preventing efficient execution on AI processors due to shrinkage of feature maps during convolution, which limits the number of convolutions that can be performed and underutilizes node elements on the edges and corners.
Innovation Solution
The introduction of Point Grid Convolutional (PGConv) layers that convert graph-structured data to grid-structured data, employing padding techniques such as zero and circular padding to maintain feature map size, and utilize attention mechanisms to generate dynamic convolution kernels, enhancing feature learning capacity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If graph convolutional networks are used to process graph-structured data, then the ability to handle irregular data structures is improved, but feature map shrinkage occurs during convolution which limits processing efficiency
Solution Approach 1:
The patent transforms the data representation from graph-structured to grid-structured format, changing the fundamental parameter of data organization. This allows convolution operations to be performed on regular grids while preserving the ability to handle irregular graph data through the transformation process, thereby improving processing efficiency without sacrificing adaptability.
Solution Approach 2:
The patent introduces a grid-structured intermediate representation as a mediator between the original graph-structured data and the convolution operation. This intermediate grid format enables efficient convolution processing while the transformation layers maintain the relationship with the original irregular data structure, resolving the contradiction between handling irregular data and achieving efficient processing.
2Loss of information
If conventional convolution operations are applied to graph-structured data, then feature extraction is performed, but node elements on edges and corners are underutilized
Solution Approach 1:
By changing the data structure from irregular graph to regular grid, the patent ensures that all node elements including those on edges and corners have consistent and meaningful neighborhoods. This parameter change in data organization allows uniform application of convolution kernels across all positions, eliminating the underutilization problem of boundary nodes while maintaining full feature extraction capability.
3Device complexity
If feature map size shrinks during convolution, then computational complexity is reduced, but the number of convolutions that can be performed is limited
Solution Approach 1:
The patent applies padding operations as a preliminary action before convolution to maintain feature map size. By pre-expanding the feature map boundaries with padding, the convolution operation can be performed multiple times without progressive shrinkage, enabling deeper network architectures and more convolutions to be executed while keeping computational complexity manageable through efficient padding implementations.
Data Source
AI summary
A PGConv layer extract features from grid-structured data samples. The PGConv layer may receive an input feature map including a grid representation of an object, which is generated from a graph representation of the object. The grid representation includes node elements that are arranged in a grid pattern. The PGConv layer may perform padding on the grid representation to generate an IFM that includes the node elements and the additional node elements. An additional node element may have a value of zero or a value of a node element in the grid representation. The PGConv layer may also generate an attentive kernel that includes attentive weights determined based on the IFM. The PGConv layer may generate a dynamic kernel based on the attentive kernel and a convolutional kernel generated through training. The PGConv layer may further perform MAC operations on the IFM and the dynamic kernel and generate an OFM.