Neural Network Convolution Kernel Grouping and Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks require significant computational resources for convolutions, which account for up to 90% of runtime, power consumption, and area usage, necessitating a reduction in computational complexity.
Innovation Solution
Identifying groups of similar convolution kernels and replacing them with a base kernel and scaling factors, allowing for the generation of a composite input feature map that can be processed using a single base convolution kernel, thereby reducing the number of multiply-accumulate operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple convolution kernels are used in a convolution stage, then the neural network can process diverse features effectively, but the computational complexity and runtime increase significantly
Solution Approach 1:
The patent merges multiple similar convolution kernels into a single base kernel by grouping them based on similarity metrics. Instead of applying multiple distinct convolution kernels to different input feature maps, the system combines their operations into one convolution operation with a base kernel, thereby reducing computational complexity while maintaining the ability to extract diverse features through the grouping mechanism
Solution Approach 2:
The patent changes the parameters of the convolution operation by introducing scaling factors that are applied to the input feature maps before convolution. By modifying the input features through scaling rather than using multiple different kernels, the system achieves parameter efficiency reduction while preserving the functional capability to process various feature types
2Reliability
If multiple convolution kernels are applied to process input feature maps, then feature diversity is maintained, but power consumption increases
Solution Approach 1:
The patent combines multiple convolution kernel operations into a single convolution operation. By merging the computational tasks into one operation with a base kernel and scaling factors, the system reduces the total number of multiply-accumulate operations required, directly lowering power consumption while maintaining feature processing accuracy through the grouping of similar kernels
3Reliability
If multiple convolution kernels are used, then the neural network can capture complex patterns, but execution time increases
Solution Approach 1:
The patent merges multiple sequential convolution operations into a single convolution operation. By grouping similar convolution kernels and combining their effects through scaling factors applied to the input feature maps, the system reduces the number of operations that must be executed, thereby decreasing execution time while preserving the pattern recognition capability through the maintained functional equivalence
4Productivity
If a single base convolution kernel is used with scaling factors, then computational operations are reduced, but the complexity of determining groups and scaling factors increases
Solution Approach 1:
The patent performs the grouping of convolution kernels and determination of scaling factors as a preliminary action during the network setup or training phase. By pre-computing which kernels belong together and what scaling factors to use, the system avoids the need to perform these complex determinations during actual inference time, thus maintaining high computational efficiency while managing the complexity through offline processing
Data Source
AI summary
Reducing computations in a neural network may include determining a group including a plurality of convolution kernels of a convolution stage of a neural network. The convolution kernels of the group are similar to one another. A base convolution kernel for the group may be determined. Scaling factors for a plurality of input feature maps processed by the group may be calculated. The convolution stage of the neural network may be modified to calculate a composite input feature map using the scaling factors and apply the base convolution kernel to the composite input feature map.