Neural Network Operator Segmentation for Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models in electronic devices require large amounts of memory to process and store input and output data, leading to slow image detection due to the huge amounts of data involved, especially when numerous operators are present.
Innovation Solution
The method involves determining operator subsets within the neural network model by mapping it to a singly-linked list, calculating output and input numbers for each node, establishing data pairs, and identifying node subsets to reduce memory usage and speed up image detection by releasing unnecessary data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all input and output data of operators are recorded and stored in memory, then complete image detection can be performed, but memory occupation increases and detection speed decreases
Solution Approach 1:
The patent segments the neural network operators into different groups based on their data dependency relationships. By dividing operators into segments that can be processed independently, the system can release intermediate data after each segment completes, reducing memory occupation while maintaining detection completeness and improving processing speed.
Solution Approach 2:
The patent implements a mechanism where intermediate operation results are discarded after being used by subsequent operators, and only necessary final results are retained. This selective discarding of temporary data reduces memory pressure during the detection process while ensuring that essential information is preserved for complete image detection.
2Reliability
If all input and output data of operators are stored in memory, then accurate image detection is achieved, but memory occupation becomes huge
Solution Approach 1:
The patent segments operators into groups with clear data flow boundaries, allowing intermediate results to be released after each segment. This segmentation maintains detection accuracy by preserving necessary data relationships while reducing the quantity of data simultaneously stored in memory.
Solution Approach 2:
The patent extracts and identifies redundant intermediate operation results that can be safely removed from memory storage. By taking out unnecessary data that is no longer needed for subsequent computations, the system reduces memory occupation while maintaining the accuracy required for proper image detection.
Data Source
AI summary
An image detection method obtains a neural network model including n operators. End operators are determined from the n operators. An image is input into the neural network model. Whether to delete operation results of the n operators is determined according to the n operators and the end operators when the neural network model processes the image using the n operators. A detection result of the image is output according to the operation results. The method can improve an efficiency of image detection.