Convolutional Neural Network Channel Width Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) face challenges with increased model size due to numerous convolution filters, leading to longer training times, higher data storage requirements, and increased computation time for object recognition, while maintaining precision in image-related tasks.
Innovation Solution
The method involves determining receptive field widths and reducing channel widths of convolution layers in a CNN, forming a second model structure with reduced channel widths and training it to maintain object recognition precision while reducing model size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If numerous convolution filters are used in CNN, then object recognition precision is improved, but model size increases
Solution Approach 1:
The patent changes the parameter of channel width from its original value to a reduced value, thereby reducing model size while maintaining recognition precision through compensatory mechanisms
Solution Approach 2:
The patent shifts the focus from increasing precision through more filters (one dimension) to achieving precision through increased receptive field coverage (another dimension), effectively changing the dimension of optimization
2Measurement precision
If numerous convolution filters are used in CNN, then object recognition precision is improved, but training time increases
Solution Approach 1:
The patent reduces the parameter of channel width to decrease the number of trainable parameters, directly reducing training time while maintaining precision through architectural adjustments
Solution Approach 2:
The patent trades off filter quantity (one dimension) for increased receptive field coverage (another dimension), achieving precision through a different computational pathway that requires less training time
3Measurement precision
If numerous convolution filters are used in CNN, then object recognition precision is improved, but computation time increases
Solution Approach 1:
The patent reduces channel width parameter to decrease the number of computations required per forward pass, reducing inference computation time while maintaining accuracy
Solution Approach 2:
The patent achieves precision not through more filters but through increased spatial coverage via larger receptive fields, reducing the computational burden of filter applications
4Measurement precision
If numerous convolution filters are used in CNN, then object recognition precision is improved, but data storage requirements increase
Solution Approach 1:
The patent reduces the channel width parameter, directly decreasing the storage required for filter weights and model parameters
Solution Approach 2:
The patent extracts and removes redundant channel dimensions from the network architecture, eliminating unnecessary storage requirements while preserving essential feature extraction capabilities
Data Source
AI summary
A method for adjusting a convolutional neural network includes following operations. The convolutional neural network includes convolution layers in a sequential order. Receptive field widths of the convolution layers in a first model of the convolutional neural network are determined. Channel widths of the convolution layers in the first model are reduced into reduced channel widths according to the receptive field widths of the convolution layers and an input image width. A structure of a second model of the convolutional neural network is formed according to the reduced channel widths. The second model of the convolutional neural network is trained according the structure of the second model.