Quadtree Tensor Compression for Sparse Neural Network Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks generate a high ratio of zero-value data due to computational characteristics, leading to inefficiencies in processing, storage, and compression, which existing compression methods fail to effectively address.
Innovation Solution
A neural tensor compressor using a quadtree method to identify and compress non-zero cells within tensors, determining a compression mode based on extracted parameters, and generating a bitstream for efficient data representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing compression methods are used on tensors with high ratio of zero-value data, then compression is performed, but compression efficiency is poor
Solution Approach 1:
The patent extracts and removes zero-value data from tensors before compression, separating the useful non-zero data from the redundant zero data. This extraction process improves compression efficiency by focusing computational resources only on the meaningful data portions, directly addressing the poor compression efficiency of existing methods when applied to tensors with high zero-value ratios.
Solution Approach 2:
The patent changes the parameter representation of tensor data by identifying and removing zero-value elements, transforming the data structure from a dense format to a sparse format. This parameter change enables more efficient compression by reducing the quantity of data that needs to be processed and stored, thereby improving compression efficiency while maintaining data integrity.
2Speed
If tensors with high ratio of zero-value cells are processed using conventional methods, then processing is performed, but processing speed is reduced
Solution Approach 1:
The patent extracts zero-value data from tensors and separates it from the non-zero data, allowing processing to focus only on the meaningful portions. This extraction reduces the computational burden and increases processing speed by eliminating redundant operations on zero-value elements, directly addressing the reduced processing speed caused by high zero-value ratios.
Solution Approach 2:
The patent segments the tensor data into zero-value portions and non-zero-value portions, processing each segment differently. This segmentation allows the system to apply optimized processing paths for sparse data, reducing overall processing time and improving speed by avoiding unnecessary processing of zero-value regions.
3Quantity of substance
If tensors with high ratio of zero-value data are stored, then storage is performed, but storage efficiency is low
Solution Approach 1:
The patent extracts and removes zero-value data from tensors before storage, significantly reducing the quantity of data that occupies storage capacity. This extraction eliminates wasted storage space by only storing meaningful non-zero data, directly addressing the low storage efficiency caused by high zero-value ratios in conventional storage methods.
Solution Approach 2:
The patent changes the storage parameter from dense tensor format to sparse tensor format by removing zero-value elements. This parameter change optimizes storage capacity utilization by storing only the necessary data, reducing wasted storage space while maintaining the ability to reconstruct the original tensor when needed.
4Loss of information
If existing compression methods are applied to neural network tensors, then compression is performed, but data loss increases
Solution Approach 1:
The patent extracts zero-value data which contains no meaningful information, and separately identifies the non-zero data that contains all the meaningful information. This extraction approach minimizes data loss by ensuring that all non-zero values are preserved during compression, directly addressing the increased data loss problem of existing methods when applied to neural network tensors.
Solution Approach 2:
The patent changes the compression approach by transforming the tensor from dense to sparse format, preserving all non-zero data points while removing only the zero-value placeholders. This parameter change maintains data integrity by ensuring that no meaningful information is lost during compression, while still achieving compression through the removal of redundant zero data.
Data Source
AI summary
A device configured to compress a tensor including a plurality of cells includes: a quadtree generator configured to generate a quadtree searching for a non-zero cell included in the tensor and extract at least one parameter value from the quadtree; a mode selector configured to determine a compression mode based on the at least one parameter; and a bitstream generator configured to generate a bitstream by compressing the tensor based on the compression mode.