Sparse Data Structure Compression with Mode Search and Significance Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional delta compression methods are ineffective for sparse data structures, such as sparse matrices in machine learning, due to large differences between neighboring numerical values, leading to unsatisfactory compression results.
Innovation Solution
A machine learning compression method utilizing mode search approximation, which identifies the most repeated value (mode) and encodes it with a significance map, operating in parallel with delta compression, to efficiently compress sparse data structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional delta compression is used on sparse data structures, then compression is attempted, but compression effectiveness deteriorates due to large differences between neighboring numerical values
Solution Approach 1:
The patent changes the compression parameter from delta encoding (difference between neighboring values) to mode-based encoding (frequency of most common value). This parameter transformation makes the compression effective for sparse data structures where many values repeat, rather than relying on smooth gradients between neighboring elements.
Solution Approach 2:
Instead of encoding the difference from one value to the next (conventional approach), the patent inverts the approach by encoding the frequency distribution of values. The significance map inverts the traditional compression paradigm by storing which positions have non-zero values rather than storing the actual difference values, which works better for sparse data with large variations.
2Reliability
If mode search approximation is implemented, then compression effectiveness improves, but computational complexity increases
Solution Approach 1:
The patent applies partial action by performing mode search only on subsets of the data or using approximate mode finding rather than exhaustive search. The ternary tree processing divides the data into manageable segments, performing mode analysis on portions rather than the entire dataset at once, reducing computational complexity while maintaining compression effectiveness.
Solution Approach 2:
The patent segments the data processing through ternary tree structure, dividing the sparse data into hierarchical segments. This segmentation allows mode search to be performed on smaller subsets independently, reducing the overall computational complexity from O(n) exhaustive search to a more efficient segmented approach while maintaining accuracy.
3Productivity
If ternary tree processing is used for mode identification, then processing speed improves, but structural complexity increases
Solution Approach 1:
The patent introduces a hierarchical dimensional structure using ternary trees, transforming the flat data structure into a multi-level tree structure. This dimensional change allows parallel processing at different tree levels, improving processing speed through hierarchical decomposition. The ternary structure (three branches per node) provides better parallelization opportunities compared to binary trees.
Solution Approach 2:
The ternary tree structure performs preliminary organization of data before the actual mode search. By pre-building the hierarchical structure and grouping similar values together in the tree, the subsequent mode identification operates on already-organized data, improving processing speed. The preliminary structuring enables faster traversal and comparison operations.
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.