Sparse Data Decompression Using SCATTER for Neural Network Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional data decompression techniques for compressed sparse data are complex and require costly arithmetic units, making them inefficient for processing in deep learning networks and other applications.
Innovation Solution
A method and system for decompressing compressed data by utilizing an N element data structure and associated metadata to reconstruct an M element data structure, with optional partial decompression to generate less densely compressed data, using instructions like SCATTER to write values from the N element data structure to the M element data structure based on metadata specifying their original locations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional data decompression techniques are used, then compressed data can be decompressed to original format, but the decompression process becomes complex and requires costly arithmetic units
Solution Approach 1:
The decompression process is segmented into multiple stages: partial decompression to an intermediate format, processing in the intermediate format, and optional final decompression. This segmentation allows the system to avoid the complexity of traditional full decompression while maintaining the ability to process data effectively.
Solution Approach 2:
The patent applies partial decompression by decompressing only a portion of the compressed data to an intermediate format rather than fully decompressing to the original format. This partial action reduces the computational complexity and resource requirements while still enabling effective data processing for machine learning operations.
2Quantity of substance
If compressed data is used directly, then storage and bandwidth resources are saved, but hardware units cannot process the compressed format
Solution Approach 1:
The system performs partial decompression to an intermediate format that is sufficiently decompressed for hardware processing while maintaining better compression than the original uncompressed format. This allows hardware units to process the data in the intermediate format without requiring full decompression to the original format.
Solution Approach 2:
The patent changes the compression parameter by introducing an intermediate compression ratio between the highly compressed original format and the fully uncompressed format. This intermediate parameter enables hardware processing compatibility while maintaining efficient storage and bandwidth utilization.
3Ease of operation
If full decompression is performed, then data is available in original format for processing, but computational resources and time are increased
Solution Approach 1:
The patent performs partial decompression to an intermediate format rather than full decompression to the original format. This partial action provides data in a sufficiently processed state for machine learning operations while significantly reducing the computational time and resources required compared to full decompression.
Solution Approach 2:
The intermediate decompression format serves as a preliminary processing step that prepares data for machine learning operations without completing the full decompression process. This preliminary action enables efficient processing by performing necessary transformations before the actual machine learning computation begins.
Data Source
AI summary
Compressed data is oftentimes beneficial for reducing the computing resources required, for example, to transmit and store data. The compression of data is particularly useful when dealing with sparse data (data that includes numerous zeros or near-zero values) and only non-zero values above a certain threshold have significance. When dealing with compressed data, oftentimes the data needs to be decompressed for processing (e.g., by deep learning networks or other applications configured to operate on sparse, or other uncompressed data). Instructions are disclosed for supporting the decompression of compressed data by a processing unit such as a CPU and GPU.