GCN Accelerator Architecture for Faster Graph Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graphs, particularly unstructured graphs, are difficult for machines to process due to variations in node and edge labeling, ordering, and weight assignments, making it challenging to compare similarity between graphs.
Innovation Solution
An accelerator is employed to extract features from graphs using Graph Convolutional Networks (GCNs), incorporating a multiplication module and an accumulate and aggregate (ACG) module, with components like a pre-fetcher, pruner, and pre-processor to optimize data processing and minimize memory access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine processing methods are used on graphs, then the processing complexity increases due to variations in node and edge labeling, ordering, and weight assignments, but the ability to determine graph similarity remains insufficient
Solution Approach 1:
The patent transforms graph data into numerical feature vectors through Graph Convolutional Network layers, changing the representation parameters from variable labels and weights to standardized numerical features. This allows similarity determination through numerical comparison while reducing processing complexity.
Solution Approach 2:
The accelerator introduces an intermediary feature extraction process between the raw graph data and similarity comparison. The GCN layers act as mediators that convert complex graph structures into comparable feature vectors, simplifying the final similarity determination step.
2Productivity
If feature extraction from graphs is performed using conventional methods, then the processing time increases due to frequent memory access, but the feature extraction capability is needed for deep learning applications
Solution Approach 1:
The patent segments the feature extraction process into distinct functional modules (multiplication module, ACG module, pre-fetcher, pruner) that can operate in parallel. This segmentation enables simultaneous data processing and memory management, reducing overall memory access time while maintaining extraction capability.
Solution Approach 2:
The pre-fetcher component performs preliminary action by loading required data into the accelerator before it is needed for processing. This advance data preparation reduces actual memory access time during the critical feature extraction computation, improving productivity.
3Productivity
If graph data is processed with all variations in labeling and weighting, then the data completeness is maintained, but the processing efficiency decreases due to unnecessary computations
Solution Approach 1:
The pruner component extracts and removes zero-value elements from the graph data before processing. By taking out these unnecessary zero elements, the system reduces computation overhead and improves processing efficiency while preserving all meaningful graph information that contributes to feature extraction.
Data Source
AI summary
A device is disclosed. The device may include a multiplication module, an accumulation and aggregation (ACG) module, and a control unit. The multiplication module may perform a multiplication based on a node data for a graph or a weight data. The ACG module may perform an accumulation and aggregation operation based on the multiplication module or an edge data for the graph. The control unit may manage the multiplication module and the ACG module.