Processing Graph Sectioning for Tiled Large-Image Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Processing large-sized images in machine learning models, such as Convolutional Neural Networks, is computationally inefficient and often results in loss of information or requires complex patch-based approaches that are time-consuming and expensive.
Innovation Solution
Implementing tiling techniques in neural networks to divide input tensors into non-overlapping or overlapping tiles, with optional zero-padding, to optimize memory usage and processing efficiency, using reconfigurable data processors like GPUs, FPGAs, and CGRAs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large-sized images are processed directly in machine learning models, then processing accuracy is maintained, but computational efficiency deteriorates and processing time increases
Solution Approach 1:
The patent divides large input images into multiple smaller tiles or patches that can be processed independently and in parallel. This segmentation allows the system to maintain processing accuracy on each tile while significantly improving computational efficiency through parallel processing and reduced memory requirements.
2Productivity
If tiling techniques are implemented to improve processing efficiency, then computational efficiency improves, but information loss may occur at tile boundaries
Solution Approach 1:
The patent introduces padding regions at the boundaries of tiles, which act as intermediaries to preserve information that would otherwise be lost at tile boundaries. These padded regions contain border pixels from adjacent tiles, ensuring that convolution operations at tile edges have access to the necessary contextual information.
Solution Approach 2:
The patent performs preliminary actions by pre-computing and storing border pixel information in padded regions before the main processing occurs. This preliminary preparation ensures that when tiles are processed independently, the boundary information is already available, preventing information loss without requiring complex post-processing.
3Loss of information
If patch-based approaches are used to process large images, then information loss is reduced, but processing time and computational cost increase
Solution Approach 1:
The patent merges the advantages of different approaches by combining tiling with intelligent padding strategies. This allows the system to process images in efficient tiles while the padding mechanism ensures that boundary information is preserved, achieving both speed and accuracy without the computational overhead of traditional patch-based methods.
Data Source
AI summary
Disclosed is a computer-implemented method that includes sectioning a processing graph for an application into a sequence of sections, the sequence of sections including at least a first section followed by a second section. The first section is configured to generate a first output. The second section is configured to generate a second output.