CNN Stride and Dilation Tuning for Lower Multiplication Load
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks require a large number of multiplications, which can overwhelm slow devices and necessitate high-performance hardware, highlighting the need for a method to reduce computational load.
Innovation Solution
A parameter optimization device and method that optimizes convolutional neural network structure by adjusting stride and dilation parameters in convolution layers, reducing unnecessary calculations through layer detection and modification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If convolutional neural network operations are performed with standard parameters, then recognition accuracy is maintained, but computational load becomes excessive for resource-constrained devices
Solution Approach 1:
The patent applies parameter changes by modifying stride and dilation values in convolutional layers. Specifically, it changes the stride parameter to skip certain positions during convolution, and adjusts dilation to control the spacing of kernel elements, thereby reducing the number of multiplications while preserving essential feature extraction capabilities.
Solution Approach 2:
The patent extracts and removes redundant calculations by identifying and eliminating unnecessary convolution operations. Through selective application of stride and dilation modifications, it takes out computationally expensive operations that contribute minimally to recognition accuracy, particularly on resource-constrained devices.
2Productivity
If device performance is limited, then hardware cost is reduced, but convolutional neural network operation speed decreases
Solution Approach 1:
The patent uses parameter changes to optimize convolution operations for resource-constrained devices. By adjusting stride and dilation parameters, it reduces computational complexity without requiring advanced hardware, enabling faster operations on simpler devices while maintaining acceptable recognition performance.
3Measurement precision
If kernel size and convolution range are increased, then feature extraction capability is improved, but number of multiplications increases significantly
Solution Approach 1:
The patent applies parameter changes by modifying dilation values to control the effective convolution range. Increased dilation creates spacing between kernel elements, allowing broader feature extraction coverage without proportionally increasing the number of multiplications, thus improving feature extraction capability while controlling computational complexity.
Solution Approach 2:
The patent segments the convolution operation by using stride to skip positions and dilation to space out kernel elements. This segmentation divides the convolution process into non-contiguous operations, reducing the total number of multiplications while maintaining feature extraction capability across different spatial regions.
Data Source
AI summary
A parameter optimization device 800 optimizes input CNN structure information and outputs optimized CNN structure information, and includes stride and dilation use layer detection means 811 for extracting stride and dilation parameter information for each convolution layer from the input CNN structure information, and stride and dilation use position modification means 812 for changing the stride and dilation parameter information of the convolution layer.