Neural Network Filter Optimization for Electronic Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks, particularly CNNs, require a large number of operations and high-capacity parameters, leading to degradation in operation speed and inefficient memory usage due to the repetitive convolution operations and extensive filtering processes.
Innovation Solution
The solution involves preprocessing and optimizing neural network filters by identifying and storing operation instructions for 'main patterns' that occupy a significant operation ratio, removing zero elements, and using optimization data with non-zero elements and identification information to reduce the computational burden and memory requirements, allowing for faster and more efficient filtering operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolution operations are repeated for feature map generation in conventional CNN, then filtering accuracy is improved, but operation amount increases and operation speed degrades
Solution Approach 1:
The patent applies preliminary action by pre-processing filters before convolution operations to identify and extract main patterns. This preparation work is done in advance, so that during actual inference, only the essential pattern matching needs to be performed, significantly reducing the real-time computational burden while preserving filtering accuracy.
Solution Approach 2:
The patent extracts main patterns from convolution filters by identifying the most significant weight configurations. Instead of performing complete convolution operations with all filter weights, the system extracts and retains only the dominant pattern structures, eliminating redundant computations while maintaining the essential filtering capability.
2Measurement precision
If high-capacity neural network parameters are used for comprehensive filtering, then model accuracy is improved, but memory usage increases and efficiency decreases
Solution Approach 1:
The patent extracts and retains only the essential pattern information from full convolution filters. By identifying main patterns that capture the most important weight configurations, the system removes redundant parameter storage requirements while preserving the critical information needed for accurate filtering operations.
Solution Approach 2:
The patent transforms the parameter representation from complete convolution kernel weights to compressed main pattern descriptors. This parameter transformation reduces the storage requirements from O(k×k×c) for a full kernel to a much smaller representation that captures only the dominant pattern structures, thereby reducing memory usage while maintaining model accuracy.
3Measurement precision
If a plurality of filters are applied to input data for feature extraction, then recognition accuracy is improved, but operation amount increases and memory efficiency decreases
Solution Approach 1:
The patent merges multiple filter operations by identifying and extracting common main patterns across different filters. Instead of independently processing each filter, the system combines redundant pattern matching operations into unified main pattern extractions, reducing the overall operational complexity while maintaining comprehensive feature extraction capability.
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
An electronic device is provided. The electronic device includes a memory and a processor configured to check an operation instruction for filtering input data of a neural network for each filter of a main pattern selected from a plurality of filters generated according to learning by the neural network, and store the checked operation instruction in the memory.