Point Grid Convolution for Efficient Graph Data Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph convolutional networks (GCNs) face inefficiencies due to irregular workloads when processing graph-structured data, limiting their performance on AI processors like GPUs and CPUs, and preventing efficient execution.
Innovation Solution
Converting graph-structured data to grid-structured data through semantic grid transformation, allowing for point grid convolution that preserves intrinsic relationships and enables more efficient convolutional operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If graph convolutional networks process graph-structured data directly, then they can preserve intrinsic relationships, but they suffer from irregular workloads that limit efficiency on AI processors
Solution Approach 1:
The patent segments graph-structured data into multiple grids based on node attributes and spatial relationships. Each grid captures local structural patterns, allowing convolutional operations to be performed on regular grid structures while preserving the intrinsic relationships of the original graph through the segmented representation.
Solution Approach 2:
The patent introduces grids as an intermediary structure between the original graph and the convolutional network. The graph is transformed into grid-structured data that serves as a mediator, enabling efficient convolutional operations while the grid structure itself preserves the essential relationships through its construction from graph nodes and edges.
2Measurement precision
If graph convolutional networks use irregular graph structures, then they maintain data fidelity, but they prevent efficient execution on GPUs and CPUs
Solution Approach 1:
The patent changes the structural parameters of the data representation by transforming irregular graph structures into regular grid structures. This parameter change involves reorganizing node and edge information into grid cells based on spatial coordinates and attribute values, maintaining data fidelity while enabling processor-efficient operations.
Solution Approach 2:
The patent introduces a grid dimension to organize graph data, transforming the traditional graph structure into a multi-dimensional grid representation. This dimensional change allows data to be accessed and processed in a manner compatible with GPU and CPU memory structures, improving execution efficiency without losing the essential relationships.
3Speed
If conventional convolutional operations are applied to graph data, then processing speed improves, but the irregular structure prevents effective convolution
Solution Approach 1:
The patent segments the graph into grid-based regions, allowing conventional convolutional operations to be applied to each grid segment independently. This segmentation enables the use of optimized convolution kernels that run at high speed on standard hardware while the segmentation process itself handles the irregularity of the original graph structure.
Data Source
AI summary
A graphical representation of an object (e.g., a 2D image) is transformed to a grid representation of the object. The grid representation adopts a structure of a grid. Graph nodes are extracted from the graphical representation and arranged based on the structure. An anchor node may be selected from the graph nodes and assigned to an element of the grid. Other graph nodes can be assigned to other elements of the grid based on their relationships with the anchor node. The grid representation can be processed by a CNN including one or more convolutional layers. A convolutional layer may receive the grid representation, generates variants of the grid representations, and extract features based on the variants. The output of the CNN can be used to determine a condition of the object, e.g., to generate a 3D graphical representation of the object that shows a pose of the object.