Subgraph Tile Fusion for Convolutional Neural Network Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks face inefficiencies due to complex data flows and sequential processing of multiple layers, which can be simplified by partitioning the network into independently run subgraphs that can be executed in parallel.
Innovation Solution
The method involves partitioning the network into subgraph nodes, determining the layer order, input, weight, and output layers, and fusing these components to enable parallel execution, reducing complexity and improving data flow efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If convolutional neural networks process multiple layers sequentially, then each layer can be computed with standard convolution operations, but the overall processing efficiency is reduced due to sequential execution
Solution Approach 1:
The patent partitions the convolutional neural network into multiple independent subgraphs, where each subgraph represents a separable portion of the network that can be processed independently. This segmentation allows parallel execution of subgraphs while maintaining the computational integrity of the original sequential network, thereby improving processing efficiency without requiring complete restructuring of the network architecture.
2Productivity
If the network is partitioned into subgraph nodes, then parallel processing can be enabled, but the complexity of partitioning and determining layer order increases
Solution Approach 1:
The patent performs preliminary analysis to determine the execution order of subgraphs before actual processing begins. By pre-determining the layer order and identifying independent subgraphs that can be processed in parallel, the system avoids complex runtime decision-making and reduces the operational complexity of parallel processing while maintaining the ability to execute multiple subgraphs concurrently.
3Productivity
If subgraphs are fused together, then the number of operations can be reduced and efficiency improved, but the complexity of fusing multiple layers increases
Solution Approach 1:
The patent merges multiple convolutional layers into fused operations within subgraphs, combining several sequential operations into a single integrated computational unit. This merging reduces the total number of operations and memory accesses by eliminating redundant intermediate steps, while the fusion is performed within bounded subgraphs to control the complexity of the merging process.
Data Source
AI summary
A method of subgraph tile fusion in a convolutional neural network, including partitioning a network into at least one subgraph node, determining a layer order of at least one layer of the at least one subgraph node, determining a input layer of the at least one subgraph node, determining a weight layer of the at least one subgraph node, determining a output layer of the at least one subgraph node and fusing the at least one subgraph node, the input layer, the weight layer and the output layer in the layer order.