Winograd Convolution for CNN Resource Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) require significant computational resources and power due to the large number of multiplications involved in conventional convolution operations, leading to high latency and resource utilization issues.
Innovation Solution
The implementation of a Winograd transform-based method for convolution, which reduces the number of multiplications by transforming and element-wise multiplying image patches and filter kernels, and then inverse transforming the results, allowing for fewer resource-intensive computations while maintaining high numerical stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional convolution operations are used in CNNs, then accurate image processing and classification can be achieved, but the number of multiplications and computational resources required increases significantly
Solution Approach 1:
The patent transforms the convolution operation parameters by applying Winograd transform to change the computational form from standard convolution to a transformed domain operation. This parameter transformation reduces the number of multiplications required while maintaining the same output accuracy, directly resolving the contradiction between precision and device complexity
Solution Approach 2:
The patent substitutes the mechanical multiplication operation with an equivalent transformation-based computation. By replacing direct convolution multiplications with Winograd transform operations followed by element-wise multiplications and inverse transforms, the system achieves the same mathematical result with fewer computational resources
2Productivity
If conventional convolution operations are used in CNNs, then complete processing of large images can be performed, but power consumption increases due to the massive amount of computational resources required
Solution Approach 1:
The Winograd transform changes the computational parameters of the convolution operation, transforming it into a form that requires fewer multiplications. This parameter transformation maintains complete image processing capability while reducing the energy-consuming multiplication operations, thereby resolving the contradiction between productivity and power consumption
3Measurement precision
If conventional convolution operations are used in CNNs, then thorough feature extraction can be achieved, but latency increases due to the large number of multiplications involved
Solution Approach 1:
The patent substitutes the time-consuming multiplication-heavy convolution operation with a transformed computation approach. By replacing sequential multiplications with Winograd transform operations that enable more efficient computation patterns, the system maintains feature extraction quality while reducing processing latency
Data Source
AI summary
Method and system relating generally to convolution is disclosed. In such a method, an image patch is selected from input data for a first channel of a plurality of input channels of an input layer. The selected image patch is transformed to obtain a transformed image patch. The transformed image patch is stored. Stored is a plurality of predetermined transformed filter kernels. A stored transformed filter kernel of the plurality of stored predetermined transformed filter kernels is element-wise multiplied by multipliers with the stored transformed image patch for a second channel of the plurality of input channels different from the first channel to obtain a product. The product is inverse transformed to obtain a filtered patch for the image patch.