Neural Network Channel Displacement With 1×1 Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural networks (DNNs) face inefficiencies due to the requirement of multiple convolution operations, particularly in lightweight structures like MobileNet and ShuffleNet, which reduce operation efficiency.
Innovation Solution
Displace channels of an input feature map according to a preset displacement parameter and perform convolution with a 1×1 kernel to obtain an output feature map, using a combination of displacement and 1×1 convolution to aggregate information on both channel and spatial dimensions, thereby reducing the need for multiple convolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If Depthwise Convolution is used to split complex convolution operation, then calculation complexity is reduced, but operation efficiency decreases due to multiple convolution operations being required
Solution Approach 1:
The patent combines displacement operation and convolution operation into a single integrated step. Instead of performing depthwise convolution separately followed by pointwise convolution, the method displaces channel data and then performs a single 1x1 convolution to aggregate information, merging what would traditionally be separate operations into one unified process that reduces overall computational steps.
Solution Approach 2:
The patent performs displacement operation as a preliminary action before the convolution operation. By pre-displacing the channel data according to preset displacement parameters, the system prepares the feature map in advance, allowing the subsequent 1x1 convolution to directly aggregate information without requiring additional separate convolution steps that would follow traditional depthwise convolution procedures.
2Loss of information
If multiple convolution operations are used in DWConv, then information aggregation on channel dimension is achieved, but DNN operation efficiency is reduced
Solution Approach 1:
The patent merges the displacement operation and the 1x1 convolution operation into a single integrated processing step. This combination allows the system to achieve information aggregation on both channel and spatial dimensions simultaneously through one convolution operation, eliminating the need for multiple separate convolution operations that would traditionally be required in depthwise convolution approaches.
Solution Approach 2:
The patent performs displacement as a preliminary action that prepares the feature map for efficient information aggregation. By displacing channel data before the convolution operation, the system enables the subsequent 1x1 convolution to aggregate information across both channel and spatial dimensions in a single operation, rather than requiring multiple separate convolution steps.
Data Source
AI summary
A deep neural network operation method and apparatus are provided. The method comprises: obtaining an input feature map of a network layer; displacing respectively, according to a preset displacement parameter, each of channels of the input feature map of the network layer along axes, to obtain a displaced feature map, wherein the preset displacement parameter comprises displacement amounts of the channel in the axes; and performing convolution operation on the displaced feature map with a 1×1 convolution kernel to obtain an output feature map of the network layer. The operation efficiency of the DNN can be improved through the above method.