Convolution Block Architecture for Neural Network Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network processing systems face inefficiencies in cell utilization during depth-wise convolution operations, leading to suboptimal performance.
Innovation Solution
A computing system architecture is proposed, featuring a convolution block with a sequence of three convolution layers: a first layer performing 1×1 convolution, a second layer performing n×n convolution on channels, and a third layer performing another 1×1 convolution, which enhances cell utilization and processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If depth-wise convolution operation is applied to the multiply-accumulate (MAC) operation device, then the neural network can perform depth-wise convolution, but cell utilization is decreased
Solution Approach 1:
The convolution operation is divided into multiple sequential stages: first convolution layer (1×1 kernel), second convolution layer (n×n kernel), and third convolution layer (1×1 kernel). This segmentation allows each layer to perform specialized operations that collectively achieve depth-wise convolution functionality while maintaining high cell utilization in the MAC operation device
Solution Approach 2:
The patent transforms the single-stage depth-wise convolution operation into a multi-layer architecture that operates across different dimensional spaces. The 1×1 convolutions operate in the channel dimension while the n×n convolution operates in the spatial dimension, creating a dimensional transformation that resolves the utilization conflict
2Device complexity
If a single convolution layer performs depth-wise convolution, then the operation is simple, but processing performance is suboptimal
Solution Approach 1:
The first convolution layer performs preliminary processing on the input data before it enters the second convolution layer. This preliminary action prepares the data in a format that optimizes subsequent processing, enabling the multi-layer structure to achieve superior performance compared to a single layer
Solution Approach 2:
The three convolution layers are arranged in continuous sequence where the output of each layer immediately becomes the input of the next. This continuous action eliminates idle periods and ensures that the MAC operation device continuously performs useful computations, maximizing processing performance
Data Source
AI summary
Provided herein may be a computing system and method of operating the same. The computing system may include an operating component including at least one convolution block, and a controller configured to control the operating component to perform convolution operations, wherein the at least one convolution block includes a first convolution layer configured to perform a first convolution operation on input data based on a 1×1 kernel to generate first result data, a second convolution layer configured to perform second convolution operations on 10 respective channels of first result data based on an n×n kernel, where n is a natural number of 2 or greater, and sum result values of the convolution operations to generate second result data, and a third convolution layer configured to perform a third convolution operation on the second result data based on the 1×1 kernel to generate final result data.