Frequency Component Convolution for Memory-Efficient Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning systems face challenges in processing information effectively, particularly in decomposing and integrating frequency components for accurate feature extraction and memory efficiency, especially when using wavelet transforms in deep neural networks.
Innovation Solution
An information processing system that performs a discrete wavelet transform on feature quantities, decomposes the results into first and second frequency components, and applies convolution processes on each component separately before integrating them, allowing for more accurate and memory-efficient output generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If wavelet transform is applied to decompose frequency components in deep learning, then feature extraction accuracy is improved, but memory cost increases
Solution Approach 1:
The patent applies segmentation by dividing the wavelet transform process into distinct stages: decomposition into frequency components, separate convolution operations on each component, and selective integration. This segmentation allows efficient memory management by processing different frequency components independently rather than loading all data simultaneously, thus maintaining high feature extraction accuracy while reducing overall memory cost.
2Manufacturing precision
If convolution process is applied to each frequency component separately, then processing precision is improved, but device complexity increases
Solution Approach 1:
The system segments the processing task by creating dedicated processing paths for different frequency components (low-frequency and high-frequency branches). Each branch performs convolution operations independently on its assigned frequency component, which improves processing precision by treating each component appropriately. The complexity is managed through modular design where each segment is a self-contained processing unit.
Solution Approach 2:
The patent introduces a frequency dimension by decomposing the input signal into multiple frequency components before processing. This transforms a single complex processing task into multiple simpler parallel tasks operating in the frequency domain. The separate convolution operations on different frequency components achieve higher precision while the parallel structure manages complexity through dimensional expansion.
3Reliability
If multiple frequency components are processed and integrated, then information quality is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by performing wavelet decomposition and frequency component separation at the beginning of the processing pipeline, before convolution operations. This preliminary organization of data into frequency components enables subsequent parallel processing of different frequency bands, improving information quality through comprehensive multi-frequency analysis while reducing overall processing time through parallel execution.
Solution Approach 2:
The processing architecture implements periodic action by alternating between different processing stages: decomposition phase, parallel convolution phase on different frequency components, and integration phase. This periodic structure allows efficient resource utilization and enables parallel processing of multiple frequency components simultaneously, maintaining high information quality while minimizing processing time through optimized stage transitions.
Data Source
AI summary
An information processing system includes: an input unit that inputs a feature quantity; a transform unit that performs a discrete wavelet transform on the feature quantity; a decomposition unit that decomposes a result of the discrete wavelet transform into a first frequency component and a second frequency component; a first processing unit that performs a convolution process on the first frequency component; a second processing unit that performs a convolution process on the second frequency component; an integration unit that integrates a result of the convolution process for the first frequency component and a result of the convolution process for the second frequency component, thereby to generate an integrated information; and an output unit that outputs the integrated information.