Sparse Data Compression Using Mode Search Approximation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional delta compression methods fail to provide satisfactory results for sparse data structures like sparse matrices used in machine learning due to large differences between neighboring numerical values, which are not smooth and contain few gradients.
Innovation Solution
A data compression method utilizing mode search approximation, where the most repeated value (mode) is extracted and encoded with a significance map, complemented by a ternary tree processing algorithm for rapid mode approximation, and can operate in parallel with delta compression.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of substance
If conventional delta compression is used on sparse data structures, then compression can be applied to data, but compression efficiency deteriorates due to large differences between neighboring numerical values
Solution Approach 1:
The patent changes the compression parameter from delta encoding to mode-based encoding. Instead of storing differences between consecutive values, the system identifies the mode (most frequent value) in data blocks and encodes deviations from this mode, which is particularly effective for sparse data where many values repeat or are close to zero.
Solution Approach 2:
The patent implements a dynamic compression approach by using a significance map that adaptively tracks the position and value of the mode across data blocks. The compression scheme dynamically adjusts to the data characteristics by identifying and encoding only the significant deviations from the mode, rather than using a static encoding method.
2Productivity
If mode search approximation is implemented, then compression efficiency for sparse data structures is improved, but device complexity increases due to additional processing algorithms
Solution Approach 1:
The patent segments the sparse data structure into manageable blocks and applies mode search approximation independently to each block. This segmentation allows the complex mode search algorithm to operate on smaller, more manageable data subsets, reducing the overall computational burden while maintaining compression efficiency.
Solution Approach 2:
The patent implements a simplified mode search approximation that performs partial mode searching rather than exhaustive search. By using approximation techniques and significance maps that track only the most important deviations, the system achieves good compression ratios without the full computational cost of exact mode search across entire datasets.
3Speed
If ternary tree processing algorithm is used for rapid mode approximation, then processing speed is improved, but device complexity increases due to tree structure implementation
Solution Approach 1:
The patent introduces a ternary tree structure that adds a dimensional organization to the mode search process. The ternary tree divides the search space into three branches at each node, enabling parallel exploration of different mode candidates and significantly speeding up the mode approximation process compared to linear search methods.
Solution Approach 2:
The ternary tree structure performs preliminary organization of data values before the actual mode search. By pre-building the ternary tree from the data blocks, the system creates an optimized search structure that allows rapid mode identification during compression operations, avoiding repeated expensive searches.
Data Source
AI summary
Embodiments are generally directed to compression for compression for sparse data structures utilizing mode search approximation. An embodiment of an apparatus includes one or more processors including a graphics processor to process data; and a memory for storage of data, including compressed data. The one or more processors are to provide for compression of a data structure, including identification of a mode in the data structure, the data structure including a plurality of values and the mode being a most repeated value in a data structure, wherein identification of the mode includes application of a mode approximation operation, and encoding of an output vector to include the identified mode, a significance map to indicate locations at which the mode is present in the data structure, and remaining uncompressed data from the data structure.