Frequency Domain Data Compression for Neural Network Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing convolution methods using Fourier transforms for image processing result in increased memory consumption due to the large amount of data generated during the convolution operation, particularly in neural networks with multiple layers, which can lead to processing inefficiencies, especially in embedded systems with limited memory capacity.
Innovation Solution
A data compression apparatus and method that calculates frequency domain coefficients using Fourier transforms, omits coefficients below a threshold, and generates feature maps using inverse Fourier transforms, reducing the data stored and processed, thereby compressing data efficiently and minimizing memory consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If convolution is performed using Fourier transform to achieve higher computation speed, then processing speed is improved, but memory consumption increases due to the large amount of frequency domain data generated
Solution Approach 1:
The patent extracts and stores only the significant frequency components (those with magnitude above a threshold) from the full frequency domain data. By separating and retaining only the essential coefficients needed for accurate convolution results, the system reduces memory consumption while preserving the computational advantages of Fourier transform-based convolution.
Solution Approach 2:
The patent applies different storage strategies to different frequency components based on their significance. Significant coefficients (above threshold) are stored in memory for accurate processing, while insignificant coefficients (below threshold) are discarded. This localized quality approach optimizes the balance between memory usage and computational accuracy.
2Measurement precision
If all frequency domain coefficients are stored to maintain convolution accuracy, then processing accuracy is improved, but memory capacity requirements increase
Solution Approach 1:
The patent extracts only the necessary frequency components for accurate convolution by applying a magnitude threshold. This extraction process identifies and retains only those coefficients that contribute significantly to the convolution result, eliminating the need to store all coefficients and thereby reducing memory capacity requirements while maintaining accuracy.
Solution Approach 2:
The patent uses partial action by storing only a subset of frequency coefficients rather than all coefficients. By applying a threshold criterion, the system stores just enough data to achieve accurate convolution results without the excess of storing complete frequency domain data, optimizing the trade-off between accuracy and memory usage.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach effectively reduces memory consumption during convolution operations, allowing for faster processing and efficient use of memory resources, even in systems with limited capacity, without significantly affecting the accuracy of feature extraction in image recognition tasks.
Implementation Method 1
calculating, from the image data, a plurality of first coefficients corresponding to different frequencies using a Fourier transform
Implementation Method 2
generating a feature map indicating a result of applying the kernel data to the image data using the first frequency domain data, the second frequency domain data, and an inverse Fourier transform
Data Source
AI summary
A processor calculates, from image data, a plurality of first coefficients corresponding to different frequencies using a Fourier transform and stores first frequency domain data, from which first coefficients whose magnitude is below a threshold have been excluded, in a memory. The processor also calculates, from kernel data, second coefficients of frequencies corresponding to the first coefficients indicated by the first frequency domain data using a Fourier transform and stores second frequency domain data in the memory. The processor uses the first frequency domain data, the second frequency domain data, and an inverse Fourier transform to generate a feature map indicating a result of applying the kernel data to the image data.