Neural Network Blocks with Grouped Convolution for Mobile Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network architectures for mobile devices, such as MobileNets, face inefficiencies in computational burden and accuracy when implemented on constrained devices like smartphones, despite efforts to reduce computational load without sacrificing performance.
Innovation Solution
A neural network block comprising an expansion layer, a compression layer, and a grouped convolution layer, with normalization techniques using batched group normalization, replacing depth-wise convolution with grouped convolution, and optimizing layer order to improve efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If depth-wise convolution is used in MobileNet blocks, then computational burden is reduced, but manufacturing precision and computational efficiency are worsened
Solution Approach 1:
The patent segments the convolution operation into grouped convolution instead of depth-wise convolution, dividing the channels into groups that can be processed more efficiently. This segmentation allows for better utilization of computational resources while maintaining reduced complexity.
Solution Approach 2:
The patent changes the normalization approach from batch normalization to batched group normalization, altering the parameters and statistics calculation method. This parameter change improves computational efficiency by better matching the grouped convolution structure and reducing redundant calculations.
2Speed
If batch normalization is applied after each layer, then training speed is improved, but device complexity increases
Solution Approach 1:
The patent merges the normalization operations across groups of channels rather than processing each channel separately. By combining the normalization steps for multiple groups, the overall complexity is reduced while maintaining the training speed benefits of normalization.
Solution Approach 2:
The batched group normalization serves multiple functions simultaneously: it normalizes within groups, reduces computational redundancy, and works seamlessly with the grouped convolution structure. This multi-functionality reduces the need for separate optimization steps.
3Loss of information
If the number of output channels in expansion layer is increased, then feature representation is improved, but loss of energy and computational cost increase
Solution Approach 1:
The patent segments the channels into groups and processes them through grouped convolution, which reduces the computational cost compared to processing all channels uniformly. This segmentation maintains feature representation quality while reducing energy consumption.
Solution Approach 2:
The batched group normalization applies different normalization statistics to different groups of channels, allowing each group to be optimized locally. This local quality approach preserves important feature characteristics while reducing overall computational burden.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A neural network block is provided. It has at least three layers: an expansion layer, a compression layer, and a grouped convolution layer. The expansion layer has more output channels than input channels. The compression layer has more input channels than output channels. A related neural network accelerator and method of inference are also provided.