Neural Network Data Processing via Zero-Value Weight Pruning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data processing based on neural network models faces limitations such as high data storage and bandwidth requirements, and low computing performance due to the inclusion of zero-value weights in processing, which do not contribute to the final calculation results.
Innovation Solution
A data processing method and device that integrates valid data from multiple data processing channels into fewer target data objects, using position information to identify and skip zero-value weights, thereby reducing storage and bandwidth needs and improving computing performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional neural network data processing is used including zero-value weights, then data storage and bandwidth requirements are high, but computing performance is low
Solution Approach 1:
The patent extracts and removes zero-value weights from the neural network data structures before processing. By identifying and eliminating these non-contributory elements, the system reduces the quantity of data that needs to be stored and transmitted, while simultaneously improving computing performance by preventing wasted computational cycles on zero-value operations.
Solution Approach 2:
The patent discards zero-value weights that do not contribute to computation results, reducing data storage requirements. The position information of these discarded weights is preserved and used to guide the processing of remaining valid data, ensuring that computational accuracy is maintained while achieving efficiency gains.
2Quantity of substance
If zero-value weights are included in processing, then data storage requirements are high, but they do not contribute to final calculation results
Solution Approach 1:
The patent performs preliminary processing to identify and remove zero-value weights before the main computation process. Position information is recorded in advance to map the locations of valid weights, enabling the system to skip unnecessary zero-value operations during processing and focus computational resources exclusively on meaningful data.
Solution Approach 2:
The patent introduces position information as an intermediary data structure that bridges the original weight matrix and the processing logic. This intermediary contains metadata about weight locations and validity, allowing the system to efficiently navigate and process only the meaningful weights without losing any computational information.
3Productivity
If all data processing channels are processed separately, then processing completeness is maintained, but processing efficiency is low
Solution Approach 1:
The patent merges the processing of multiple data processing channels by consolidating their weight matrices into a unified structure. Valid weights from different channels are integrated while maintaining their positional relationships, allowing the system to process multiple channels more efficiently without compromising the completeness or accuracy of individual channel processing.
Data Source
AI summary
A data processing method includes obtaining at least one first target data object including first target data, the first target data in each first target data object at least including all valid data corresponding to each data processing channel, each first target data corresponding to corresponding position information, the position information indicating a position of second target data corresponding to the first target data, a number of first target data objects being less than a number of data processing channels; obtaining the corresponding second target data from to-be-processed data included in a second data object corresponding to each data processing channel based on the position information corresponding to the first target data; and performing data processing on the first target data and the corresponding second target data.