Neural Network Coefficient Matrix Compression for Small Buffer Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in efficiently compressing and processing neural network coefficients, particularly in devices with small buffer sizes, where division of processing is necessary due to limited resources.
Innovation Solution
An information processing apparatus that generates and restores divided compressed data by dividing and compressing a neural network coefficient matrix in a specific range, allowing for flexible setting in certain directions while maintaining fixed settings in the filter direction, to accommodate various shapes and sizes of coefficient matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the coefficient matrix is compressed using conventional methods, then the compression rate is improved, but the ability to divide processing in devices with small buffer sizes deteriorates
Solution Approach 1:
The coefficient matrix is divided into multiple sub-matrices along the filter direction, with each sub-matrix being compressed independently. This segmentation enables processing division for devices with small buffer sizes while maintaining high compression rates through efficient encoding of each segment.
Solution Approach 2:
The patent introduces a divided range parameter that allows flexible configuration in dimensions other than the filter direction. By changing the compression strategy from a single-dimensional approach to a multi-dimensional approach with configurable ranges, the system achieves both high compression rates and processing division capability.
2Adaptability or versatility
If the divided range is set freely in all directions, then the adaptability to various coefficient matrix shapes is improved, but the processing division capability for devices with small buffer sizes deteriorates
Solution Approach 1:
The patent applies different compression strategies to different parts of the coefficient matrix. The filter direction is handled with fixed constraints for processing division, while other directions allow flexible range settings for adaptability. This local differentiation resolves the contradiction between processing division capability and shape adaptability.
Solution Approach 2:
The system dynamically configures the divided range parameters based on the specific coefficient matrix dimensions and device capabilities. The flexible setting in non-filter directions allows adaptation to various shapes, while the fixed constraint in the filter direction maintains processing division simplicity.
3Adaptability or versatility
If the divided range is set freely in the filter direction, then the processing division capability is improved, but the compression efficiency deteriorates
Solution Approach 1:
The coefficient matrix is segmented along the filter direction into manageable portions that can be processed independently. This segmentation enables processing division while maintaining compression efficiency through optimized encoding of each segment, avoiding the need for completely flexible range settings.
Solution Approach 2:
The patent changes the parameter configuration by fixing the filter direction range while allowing flexibility in other directions. This parameter adjustment resolves the contradiction by maintaining compression efficiency through structured segmentation while still enabling processing division through the fixed divided range mechanism.
Data Source
AI summary
An information processing apparatus (1) includes a processing unit (11) that generates divided compressed data (dc) by dividing and compressing a coefficient matrix (km) of a neural network, which has dimensions in a filter direction and in other directions and is adjusted to include many zero coefficients, in a divided range designed to be unable to be set freely in the filter direction but to be able to be set freely in the other directions.