Neural Network Acceleration via Spatial to Frequency Domain Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks, particularly convolutional neural networks, face inefficiencies in completing forward operations due to high computational overhead, with convolution operations accounting for more than 80% of the operating time.
Innovation Solution
The method involves converting neural network operations into a different geometric space, such as from the spatial domain to the frequency domain using transformations like FFT, DFT, or DCT, and performing operations in this new space to accelerate processing, including optimization techniques like component selection and sparsification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If convolution operations are performed in the spatial domain using traditional methods, then the operations can be completed with standard computational approaches, but the computational overhead is excessively high and the operating speed is slow
Solution Approach 1:
The patent transforms the computational parameters by changing the domain from spatial to frequency domain using Fourier transform. This parameter transformation converts convolution operations in spatial domain into multiplication operations in frequency domain, dramatically reducing computational complexity from O(n²) to O(n log n) and achieving significant speedup in neural network operations
Solution Approach 2:
The patent replaces the traditional mechanical computational approach (direct convolution in spatial domain) with a mathematical transformation approach (Fourier transform to frequency domain). This substitution changes the fundamental mechanism from iterative spatial convolution to efficient frequency-domain multiplication, eliminating the computational bottleneck
2Measurement precision
If full precision data is maintained throughout neural network operations, then computational accuracy is preserved, but data storage overhead and processing complexity increase significantly
Solution Approach 1:
The patent applies different precision requirements to different parts of the computational process. By transforming to frequency domain, many coefficients become negligibly small and can be truncated or quantized with lower precision, while only the significant frequency components maintain high precision. This local differentiation of precision requirements reduces overall storage and processing complexity while preserving essential computational accuracy
Solution Approach 2:
The patent extracts and processes only the significant frequency components from the full data spectrum. By identifying and retaining only the dominant frequency coefficients that contribute meaningfully to the computation, the system eliminates redundant low-magnitude coefficients, thereby reducing storage overhead and processing complexity without substantially compromising accuracy
Data Source
AI summary
A computing device, comprising: a computing module, comprising one or more computing units; and a control module, comprising a computing control unit, and used for controlling shutdown of the computing unit of the computing module according to a determining condition. Also provided is a computing method. The computing device and method have the advantages of low power consumption and high flexibility, and can be combined with the upgrading mode of software, thereby further increasing the computing speed, reducing the computing amount, and reducing the computing power consumption of an accelerator.