Hybrid Non-Uniform Convolution Transform Engine for Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning convolutional neural networks (CNNs) face resource-demanding computational workloads due to intensive convolution operations, which are not efficiently addressed by existing technologies.
Innovation Solution
A system and method that selectively enable processing elements with more adders than multipliers to perform convolution operations in either a transformed domain or directly, based on input size, utilizing the Winograd transform for efficient computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional convolution computations are used in CNNs, then the convolution operations can be performed, but the computational workload becomes intensive and resource-demanding
Solution Approach 1:
The patent applies parameter changes by transforming the convolution operation from the spatial domain to the frequency domain using Fast Fourier Transform (FFT). This changes the computational parameters from direct convolution (O(n²)) to frequency domain multiplication (O(n log n)), significantly reducing computational complexity and resource consumption while maintaining computation accuracy.
2Productivity
If more adders are provided in processing elements, then the transformed domain computation can be performed efficiently, but the device complexity increases
Solution Approach 1:
The patent employs dynamics by making the processing elements reconfigurable. The same processing element can be dynamically configured to perform different operations (direct convolution or transformed domain convolution) based on the input characteristics. This allows the system to adapt its structure to the computational requirements, achieving high performance for transformed domain operations without permanently increasing device complexity.
3Adaptability or versatility
If the system supports both transformed domain and direct convolution modes, then adaptability to different input sizes is improved, but the device complexity increases
Solution Approach 1:
The patent implements universality by designing processing elements that can perform multiple functions. Each processing element is capable of executing both direct convolution operations and transformed domain convolution operations, eliminating the need for separate dedicated hardware for each mode. This multi-functional design achieves adaptability to different input sizes while controlling device complexity through resource sharing.
Data Source
AI summary
A system performs convolution operations based on an analysis of the input size. The input includes data elements and filter weights. The system includes multiple processing elements. Each processing element includes multipliers and adders, with more of the adders than the multipliers. According to at least the analysis result which indicates whether the input size matches a predetermined size, the system is operative to select a first mode or a second mode. In the first mode, a greater number of the adders than the multipliers are enabled for each processing element to multiply transformed input and to perform an inverse transformation. In the second mode, an equal number of the adders and the multipliers are enabled for each processing element to multiply-and-accumulate the input. One or more of the multipliers are shared by the first mode and the second mode.